Introduction
We investigate sharp-interface variational models for polycrystalline materials that fully respect the symmetry of the underlying lattice. The energy is defined on partitions of a domain, each subset – representing an individual grain – being endowed with an orientation modeled as an element of the quotient space O ( d ) / G O(d)/G O ( d ) / G . Here, G ⊂ O ( d ) G \subset O(d) G ⊂ O ( d ) is a finite point group describing the symmetries of a reference lattice, and the local orientation is measured relative to this reference. The total energy is given by an integral over the interfaces separating adjacent grains. The integrand, or surface energy density, depends on the orientations of the two neighboring grains and on the unit normal to the interface. To ensure that the class of piecewise-constant orientation fields is closed, we assume the energy density grows superlinearly as the misorientation angle tends to zero.
A distinctive feature of the model is that the orientation variable takes values in a manifold rather than a linear space. Classical analytical tools such as regularization, approximation, and interpolation are therefore not directly applicable, and a dedicated framework for manifold-valued fields is required.
Models of this type are widely used to study polycrystalline structures in metals and alloys, see, for instance, [13 , 8 , 26 , 37 ] and references therein. Dependence on both orientations and lattice symmetry is likewise central to related problems such as polygonal partitions with prescribed area and minimal perimeter [24 ] , and optimal quantization or location of measures [11 , 12 ] . The point-group symmetry also enters variational models of fractional defects, see for example [6 , 5 , 31 ] . Analytical aspects such as semicontinuity and relaxation – though without an explicit treatment of the symmetry group G \mathcal{G} G – have been addressed in [2 , 34 , 15 ] .
The principal objective of this paper is to develop a diffuse-interface, or phase-field, approximation of the sharp-interface model described above. Specifically, we construct a phase-field approximation for surface energies with densities that vanish as the misorientation between grains tends to zero, but do so in a superlinear manner. Our approach builds on the model introduced for cohesive fracture in [19 , 21 ] . The analysis relies on two key abstract ingredients. First, compactness properties are formulated within the framework of metric-space-valued generalizations of Sobolev spaces and spaces of functions of bounded variation, as developed in [36 , 1 ] . Second, the construction of recovery sequences is based on a lifting theorem for BV functions taking values in compact manifolds, established in [14 ] (here we provide a simpler proof valid for our specific case, see Theorem 2.12 ). These tools allow us to rigorously prove the convergence of the phase-field approximation to the sharp-interface limit.
To describe the sharp-interface model in more detail, let { E i } i \{E_i\}_i { E i } i be a partition of a domain Ω ⊂ R N \Omega\subset\mathbb{R}^{N} Ω ⊂ R N , representing the material body. Each set E i E_i E i corresponds to an individual grain with assigned orientation R i ∈ O ( d ) R_i \in\mathrm{O}(d) R i ∈ O ( d ) , yielding a piecewise-constant function β = ∑ i χ E i R i \beta= \sum_i \chi_{E_i}R_i β = ∑ i χ E i R i as the order parameter. The grain boundaries are contained in the discontinuity set J β J_\beta J β of β \beta β and the corresponding energy is given by
∫ J β Ψ ( R − , R + , ν β ) d H N − 1 = ∑ i ≠ j H N − 1 ( ∂ ∗ E i ∩ ∂ ∗ E j ) Ψ ( R i , R j , ν i j ) , (1) \int_{J_\beta} \Psi(R^{-}, R^{+}, \nu_\beta)\,\mathrm{d}\mathcal{H}^{N-1} = \sum_{i\ne j} \mathcal{H}^{N-1}(\partial^{*}E_i \cap\partial^{*}E_j)\Psi(R_i,R_j,\nu_{ij}), \tag*{(1)} ∫ J β Ψ ( R − , R + , ν β ) d H N − 1 = i = j ∑ H N − 1 ( ∂ ∗ E i ∩ ∂ ∗ E j ) Ψ ( R i , R j , ν ij ) , (1)
where Ψ : O ( d ) × O ( d ) × S N − 1 → [ 0 , ∞ ) \Psi: \mathrm{O}(d) \times\mathrm{O}(d) \times\mathbb{S}^{N-1} \to[0,\infty) Ψ : O ( d ) × O ( d ) × S N − 1 → [ 0 , ∞ ) is the surface energy density, ν β \nu_\beta ν β denotes the measure-theoretic normal to J β J_\beta J β , and ν i j \nu_{ij} ν ij is the normal to the intersection of the reduced boundaries ∂ ∗ E i ∩ ∂ ∗ E j \partial^{*}E_i \cap\partial^{*}E_j ∂ ∗ E i ∩ ∂ ∗ E j . A natural function space for such energies is the space of Special functions with Bounded Variation with values in O ( d ) \mathrm{O}(d) O ( d ) and vanishing absolutely continuous part of the gradient,
S B V 0 ( Ω ; O ( d ) ) : = { β ∈ S B V ( Ω ; O ( d ) ) : ∇ β = 0 } . (2) \mathrm{SBV}_{0}(\Omega;\mathrm{O}(d)) := \{\beta\in\mathrm{SBV}(\Omega;\mathrm{O}(d)) : \nabla\beta= 0\}. \tag*{(2)} SBV 0 ( Ω ; O ( d )) := { β ∈ SBV ( Ω ; O ( d )) : ∇ β = 0 } . (2)
For the variational problems associated with (1.1) to be well-posed in this space, certain conditions on Ψ \Psi Ψ are required. In particular, lower semicontinuity of the energy follows from the BV-ellipticity of Ψ \Psi Ψ (see [3 Theorem 5.14] ). Additionally, it is essential that S B V 0 ( Ω ; O ( d ) ) \mathrm{SBV}_{0}(\Omega;\mathrm{O}(d)) SBV 0 ( Ω ; O ( d )) is sequentially closed under the relevant convergence. This requires that sequences with bounded energy cannot converge to fields with diffuse gradient. A sufficient condition is that Ψ \Psi Ψ has infinite slope at zero,
lim ∣ R + − R − ∣ → 0 Ψ ( R + , R − , ν ) ∣ R + − R − ∣ = + ∞ \lim_{|R^{+}-R^{-}|\to0}\frac{\Psi(R^{+},R^{-},\nu)}{|R^{+}-R^{-}|}=+\infty ∣ R + − R − ∣ → 0 lim ∣ R + − R − ∣ Ψ ( R + , R − , ν ) = + ∞
uniformly in ν \nu ν , see [3 Theorem 4.7] . Denoting by θ = ∣ R + − R − ∣ \theta=|R^{+}-R^{-}| θ = ∣ R + − R − ∣ the relative mismatch between neighboring grain orientations, the specific growth law θ ∣ log θ ∣ \theta|\log\theta| θ ∣ log θ ∣ of Ψ \Psi Ψ for small θ \theta θ was first identified by Read and Shockley in the ’50s [35 ] . In two dimensions, this scaling has been recently rigorously derived as the limit of a model that accounts for the elastic distortion induced by incompatibilities and crystal defects at the boundary, see [33 , 28 ] and [27 ] .
Since β \beta β is interpreted as a field taking values in O ( d ) \mathrm{O}(d) O ( d ) , invariance of Ψ \Psi Ψ under the action of G \mathcal{G} G must be enforced. In particular, boundaries between regions whose orientations agree up to multiplication by an element of G \mathcal{G} G are spurious, as they represent the same physical state, and do not contribute to the energy. Accordingly, we require Ψ ( R , G R , ν ) = 0 \Psi(R,\mathcal{G}R,\nu)=0 Ψ ( R , G R , ν ) = 0 for all R ∈ O ( d ) R\in\mathrm{O}(d) R ∈ O ( d ) , G ∈ G \mathcal{G}\in\mathcal{G} G ∈ G , ν ∈ S N − 1 \nu\in\mathbb{S}^{N-1} ν ∈ S N − 1 , and, more generally,
Ψ ( G R , G ′ R ′ , ν ) = Ψ ( R , R ′ , ν ) for all R , R ′ ∈ O ( d ) , G , G ′ ∈ G , ν ∈ S N − 1 . (3) \Psi(\mathcal{G}R,\mathcal{G}'R',\nu)=\Psi(R,R',\nu)\qquad\text{for all }R,R'\in\mathrm{O}(d),\ \mathcal{G},\mathcal{G}'\in\mathcal{G},\ \nu\in\mathbb{S}^{N-1}. \tag*{(3)} Ψ ( G R , G ′ R ′ , ν ) = Ψ ( R , R ′ , ν ) for all R , R ′ ∈ O ( d ) , G , G ′ ∈ G , ν ∈ S N − 1 . (3)
It is important to note that, even in simple examples, this invariance cannot be removed by prescribing a consistent choice of the representative for each orientation class. See Figure 1 for an illustration.
Figure 1. Sketch of a possible configuration. Left: the dots represent the local lattice orientation in a polycrystal with five grains, labeled (a), (b), (c), (d), (e). Around each interface, a possible orientation of the two basis vectors is illustrated; each interface has a change of orientation of 18 degrees. However, these changes accumulate, and if one starts with a given orientation (interface (a) − (b)) then the orientation after five interfaces differs from the initial one by 90 degrees, leading to a fictitious interface, for instance across the dashed red line in grain (a). On the right, the explicit rotation matrices are indicated (angles in degrees).
The goal of this paper is to construct variational phase-field models which converge to sharp-interface energies of the form (1) . Our approximation respects the invariance under the lattice point-group G G G , and reproduces the Read-Shockley superlinear scaling θ ∣ log θ ∣ \theta\lvert\log\theta\rvert θ ∣ log θ ∣ for small misorientation angles. The approach is motivated and inspired by the algorithms used in [8 , 26 ] for the segmentation of polycrystalline images.
Phase-field methods are widely employed in numerical simulations of sharp-interface energies, and, more in general, of free-discontinuity problems. By replacing discontinuity sets with narrow diffuse layers, they avoid the need to track interfaces explicitly and permit to use standard finite-element discretizations designed for Sobolev functions. Classical applications include image reconstruction [17 , 9 , 10 ] and fracture mechanics [29 , 25 , 16 , 30 , 38 , 32 ] .
This strategy originated with the Ambrosio-Tortorelli approximation of the Mumford-Shah functional [4 ] , in which the surface energy density depends solely on the measure of the jump set, and not on the jump amplitude. More recently, a number of phase-field approximations for cohesive fracture and segmentation energies – where the surface energy may depend on the opening – have been rigorously developed, see for instance [22 , 19 , 18 , 20 , 21 ] .
We approximate sharp-interface energies with superlinear density Ψ \Psi Ψ following the approach introduced in [19 , 21 ] , with the key additional challenge of incorporating the group invariance, which we describe in more detail below. Two further technical distinctions are worth noting: first, to recover the Read-Shockley scaling law, we adapt the construction from [19 Section 7] , [21 ] , originally designed for general superlinear power-law growth at small openings. Second, we introduce a penalization term – similar to that in [18 ] for the approximation of piecewise rigid motions with brittle interface energy – to ensure that the diffuse gradient vanishes in the limit.
Formulation of the main results
Let Ω ⊂ R N \Omega\subset\mathbb{R}^{N} Ω ⊂ R N , N ≥ 1 N \ge1 N ≥ 1 , be a bounded connected set with Lipschitz boundary, and let G ⊂ O ( d ) \mathcal{G} \subset O(d) G ⊂ O ( d ) , for d ≥ 2 d \ge2 d ≥ 2 , be a finite group, possibly with d ≠ N d \ne N d = N . We define the space of admissible order parameters that encode the lattice symmetry as
S B V G ( Ω ; M d × d ) : = { β ∈ S B V 2 ( Ω ; M d × d ) : β + ( x ) ∈ G β − ( x ) for H N − 1 -a.e. x ∈ J β } , (4) \mathrm{SBV}_{\mathcal{G}}(\Omega;\mathbb{M}^{d \times d}) := \left\{\beta\in\mathrm{SBV}^{2}(\Omega;\mathbb{M}^{d \times d}): \beta^{+}(x) \in\mathcal{G}\beta^{-}(x)\ \text{for }\mathcal{H}^{N-1}\text{-a.e. }x \in J_{\beta}\right\}, \tag*{(4)} SBV G ( Ω ; M d × d ) := { β ∈ SBV 2 ( Ω ; M d × d ) : β + ( x ) ∈ G β − ( x ) for H N − 1 -a.e. x ∈ J β } , (4)
i.e., the set of all SBV functions such that H N − 1 ( J β ) < ∞ \mathcal{H}^{N-1}(J_{\beta}) < \infty H N − 1 ( J β ) < ∞ , ∇ β ∈ L 2 \nabla\beta\in L^{2} ∇ β ∈ L 2 , and the traces on the jump set differ only by multiplication by an element of G \mathcal{G} G (we will sometimes say simply “jump in G \mathcal{G} G ” for brevity).
For ε > 0 \varepsilon> 0 ε > 0 and an open set A ⊂ Ω A \subset\Omega A ⊂ Ω we consider the energy
F ε r e d ( β , v ; A ) : = { ∫ A f ε 2 ( v ) ∣ ∇ β ∣ 2 d x + P F ε ( v ; A ) in S B V G ( Ω ; M d × d ) × H 1 ( Ω ; [ 0 , 1 ] ) , + ∞ otherwise in L 1 ( Ω ; M d × d ) × L 1 ( Ω ) , (5) \mathcal{F}^{\mathrm{red}}_{\varepsilon}(\beta,v;A) :=
\begin{cases}
\int_{A} f_{\varepsilon}^{2}(v)|\nabla\beta|^{2}\,dx+\mathrm{PF}_{\varepsilon}(v;A) & \text{in }\mathrm{SBV}_{\mathcal{G}}(\Omega;\mathbb{M}^{d \times d}) \times H^{1}(\Omega;[0,1]),\\
+\infty& \text{otherwise in }L^{1}(\Omega;\mathbb{M}^{d \times d}) \times L^{1}(\Omega),
\end{cases} \tag*{(5)} F ε red ( β , v ; A ) := { ∫ A f ε 2 ( v ) ∣∇ β ∣ 2 d x + PF ε ( v ; A ) + ∞ in SBV G ( Ω ; M d × d ) × H 1 ( Ω ; [ 0 , 1 ]) , otherwise in L 1 ( Ω ; M d × d ) × L 1 ( Ω ) , (5)
where P F ε \mathrm{PF}_{\varepsilon} PF ε is a Modica-Mortola type energy and is given by
P F ε ( v ; A ) : = ∫ A ( ( 1 − v ) 2 4 ε + ε ∣ ∇ v ∣ 2 ) d x , v ∈ H 1 ( Ω ; [ 0 , 1 ] ) . (6) \mathrm{PF}_{\varepsilon}(v;A):=\int_{A}\left(\frac{(1-v)^{2}}{4\varepsilon}+\varepsilon|\nabla v|^{2}\right)\,dx,\qquad v\in H^{1}(\Omega;[0,1]). \tag*{(6)} PF ε ( v ; A ) := ∫ A ( 4 ε ( 1 − v ) 2 + ε ∣∇ v ∣ 2 ) d x , v ∈ H 1 ( Ω ; [ 0 , 1 ]) . (6)
We notice that F ε r e d \mathcal{F}^{\mathrm{red}}_{\varepsilon} F ε red is finite on the space S B V G \mathrm{SBV}_{\mathcal{G}} SBV G , and jumps in G \mathcal{G} G of the order parameter β \beta β do not contribute to the energy. As in the models proposed in [19 , 21 ] , the choice of the prefactor f ε f_{\varepsilon} f ε in front of ∣ ∇ β ∣ 2 |\nabla\beta|^{2} ∣∇ β ∣ 2 , the so-called damage coefficient, is crucial to capture the desired behavior of the limiting surface energy density. At variance with [19 , 21 ] , and following [18 ] , we allow f ε f_{\varepsilon} f ε to take large values, enforcing that the limit fields have no diffuse gradient. More precisely, we set
f ε ( s ) : = M ε ∧ ε f ( s ) for s ∈ [ 0 , 1 [ , f ε ( 1 ) = M ε , lim ε → 0 + M ε = + ∞ , (7) f_{\varepsilon}(s):=M_{\varepsilon}\wedge\sqrt{\varepsilon}f(s)\ \text{for }s\in[0,1[,\qquad f_{\varepsilon}(1)=M_{\varepsilon},\qquad\lim_{\varepsilon\to0^{+}}M_{\varepsilon}=+\infty, \tag*{(7)} f ε ( s ) := M ε ∧ ε f ( s ) for s ∈ [ 0 , 1 [ , f ε ( 1 ) = M ε , ε → 0 + lim M ε = + ∞ , (7)
where f ∈ C ( [ 0 , 1 [ ; [ 0 , + ∞ [ ) f\in C([0,1[;[0,+\infty[) f ∈ C ([ 0 , 1 [ ; [ 0 , + ∞ [ ) is a nondecreasing function satisfying f − 1 ( 0 ) = { 0 } f^{-1}(0)=\{0\} f − 1 ( 0 ) = { 0 } , and
ℓ : = lim s → 1 − 1 − s ∣ log ( 1 − s ) ∣ f ( s ) > 0 , (8) \ell:=\lim_{s\to1^{-}}\frac{1-s}{|\log(1-s)|}f(s)>0, \tag*{(8)} ℓ := s → 1 − lim ∣ log ( 1 − s ) ∣ 1 − s f ( s ) > 0 , (8)
as for example
f ( s ) = ℓ ∣ log ( 1 − s ) ∣ 1 − s . f(s)=\ell\frac{|\log(1-s)|}{1-s}. f ( s ) = ℓ 1 − s ∣ log ( 1 − s ) ∣ .
The singularity of f f f at 1 induces the logarithmic superlinear behaviour of the surface tension at small openings (cf. Proposition 4.4), while the truncation by the diverging sequence M ε M_{\varepsilon} M ε ensures the condition ∇ β = 0 \nabla\beta=0 ∇ β = 0 in the domain of the Γ \Gamma Γ -limit (see Theorem 3.2 and Theorem 3.3).
Finally, we introduce a penalization term that asymptotically forces the limit fields to take values in O ( d ) O(d) O ( d ) . To this end, we consider an additional positive parameter δ ε → 0 + \delta_{\varepsilon}\to0^{+} δ ε → 0 + and define the modified energy
F ε ( β , v ; A ) : = F ε r e d ( β , v ; A ) + 1 δ ε ∫ A dist 2 ( β , O ( d ) ) d x . (9) \mathcal{F}_{\varepsilon}(\beta,v;A):=\mathcal{F}^{\mathrm{red}}_{\varepsilon}(\beta,v;A)+\frac{1}{\delta_{\varepsilon}}\int_{A}\operatorname{dist}^{2}(\beta,O(d))\,dx. \tag*{(9)} F ε ( β , v ; A ) := F ε red ( β , v ; A ) + δ ε 1 ∫ A dist 2 ( β , O ( d )) d x . (9)
We shall show that limiting behavior of the energies F ε \mathcal{F}_{\varepsilon} F ε depends only on the asymptotic ratio
λ : = lim ε → 0 ε δ ε ∈ [ 0 , + ∞ ] , (10) \lambda:=\lim_{\varepsilon\to0}\frac{\varepsilon}{\delta_{\varepsilon}}\in[0,+\infty], \tag*{(10)} λ := ε → 0 lim δ ε ε ∈ [ 0 , + ∞ ] , (10)
that we assume to exist. In view of the invariance of the energy F ε \mathcal{F}_{\varepsilon} F ε under the action of the symmetry group G \mathcal{G} G , it is natural to adopt the following notion of convergence modulo G \mathcal{G} G . A family ( β ε ) ε ⊂ L 1 ( Ω ; M d × d ) (\beta_{\varepsilon})_{\varepsilon} \subset L^{1}(\Omega; \mathbb{M}^{d \times d}) ( β ε ) ε ⊂ L 1 ( Ω ; M d × d ) is said to converge to β \beta β in the L G 1 L^{1}_{\mathcal{G}} L G 1 -sense if there exists a sequence G ε ∈ L ∞ ( Ω ; G ) G_{\varepsilon} \in L^{\infty}(\Omega; \mathcal{G}) G ε ∈ L ∞ ( Ω ; G ) such that
G ε β ε → β in L 1 ( Ω ; M d × d ) . (11) G_{\varepsilon}\beta_{\varepsilon} \to\beta\quad\text{in } L^{1}(\Omega; \mathbb{M}^{d \times d}). \tag*{(11)} G ε β ε → β in L 1 ( Ω ; M d × d ) . (11)
With respect to this notion of convergence, we prove a compactness result for sequences with equi-bounded energy and characterize the Γ \Gamma Γ -limit of F ε \mathcal{F}_{\varepsilon} F ε . More precisely, for any λ ∈ [ 0 , + ∞ ] \lambda\in[0,+\infty] λ ∈ [ 0 , + ∞ ] we define the sharp-interface functional
F λ ( β , v ) : = { ∫ J β g λ ( β − , β + ) d H N − 1 if β ∈ S B V 0 ( Ω ; O ( d ) ) , v = 1 , + ∞ otherwise in L 1 ( Ω ; M d × d ) × L 1 ( Ω ) , (12) \mathcal{F}^{\lambda}(\beta,v):=
\begin{cases}
\int_{J_{\beta}}g_{\lambda}(\beta^{-},\beta^{+})\,d\mathcal{H}^{N-1} & \text{if } \beta\in\mathrm{SBV}_{0}(\Omega;\mathrm{O}(d)),\ v=1,\\
+\infty& \text{otherwise in } L^{1}(\Omega;\mathbb{M}^{d\times d})\times L^{1}(\Omega),
\end{cases} \tag*{(12)} F λ ( β , v ) := { ∫ J β g λ ( β − , β + ) d H N − 1 + ∞ if β ∈ SBV 0 ( Ω ; O ( d )) , v = 1 , otherwise in L 1 ( Ω ; M d × d ) × L 1 ( Ω ) , (12)
where the surface energy density g λ g_{\lambda} g λ is determined via an asymptotic cell problem that will be introduced and analyzed in detail in § 4 .
Due to the invariance of the energy F ε \mathcal{F}_{\varepsilon} F ε under the action of G \mathcal{G} G , the limit density g λ g_{\lambda} g λ inherits the same invariance property. That is, for every R + , R − ∈ O ( d ) R^{+},R^{-}\in\mathrm{O}(d) R + , R − ∈ O ( d ) we have
g λ ( R + , R − ) = g λ ( G R + , G ′ R − ) ∀ G , G ′ ∈ G , (13) g_{\lambda}(R^{+},R^{-})=g_{\lambda}(GR^{+},G'R^{-})\quad\forall G,G'\in\mathcal{G}, \tag*{(13)} g λ ( R + , R − ) = g λ ( G R + , G ′ R − ) ∀ G , G ′ ∈ G , (13)
(see Proposition 4.1 ). In particular, the functional F λ \mathcal{F}^{\lambda} F λ depends only on the equivalence classes of β \beta β in the quotient space O ( d ) / G \mathrm{O}(d)/\mathcal{G} O ( d ) / G , namely,
F λ ( β , v ) = F λ ( G β , v ) for every ( β , v , G ) ∈ L 1 ( Ω ; M d × d ) × L 1 ( Ω ) × B V ( Ω ; G ) . (14) \mathcal{F}^{\lambda}(\beta,v)=\mathcal{F}^{\lambda}(G\beta,v)\quad\text{for every }(\beta,v,G)\in L^{1}(\Omega;\mathbb{M}^{d\times d})\times L^{1}(\Omega)\times\mathrm{BV}(\Omega;\mathcal{G}). \tag*{(14)} F λ ( β , v ) = F λ ( Gβ , v ) for every ( β , v , G ) ∈ L 1 ( Ω ; M d × d ) × L 1 ( Ω ) × BV ( Ω ; G ) . (14)
The main result of this paper is the following.
Theorem 1.1. The following properties hold:
(C) For any ( β ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d × [ 0 , 1 ] ) (\beta_{\varepsilon},v_{\varepsilon})_{\varepsilon}\subset L^{1}(\Omega;\mathbb{M}^{d\times d}\times[0,1]) ( β ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d × [ 0 , 1 ]) satisfying
lim sup ε → 0 F ε ( β ε , v ε ) < + ∞ (15) \limsup_{\varepsilon\to0}\mathcal{F}_{\varepsilon}(\beta_{\varepsilon},v_{\varepsilon})<+\infty \tag*{(15)} ε → 0 lim sup F ε ( β ε , v ε ) < + ∞ (15) there exists β ∈ S B V 0 ( Ω ; O ( d ) ) \beta\in\mathrm{SBV}_{0}(\Omega;\mathrm{O}(d)) β ∈ SBV 0 ( Ω ; O ( d )) such that, up to a subsequence,
β ε → β in L G 1 ( Ω ; M d × d ) , v ε → 1 in L 1 ( Ω ) ; (16) \beta_{\varepsilon}\to\beta\ \text{in }L^{1}_{\mathcal{G}}(\Omega;\mathbb{M}^{d\times d}),\qquad v_{\varepsilon}\to1\ \text{in }L^{1}(\Omega); \tag*{(16)} β ε → β in L G 1 ( Ω ; M d × d ) , v ε → 1 in L 1 ( Ω ) ; (16) (LB) For any ( β ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d × [ 0 , 1 ] ) (\beta_{\varepsilon},v_{\varepsilon})_{\varepsilon}\subset L^{1}(\Omega;\mathbb{M}^{d\times d}\times[0,1]) ( β ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d × [ 0 , 1 ]) and β ∈ S B V 0 ( Ω ; O ( d ) ) \beta\in\mathrm{SBV}_{0}(\Omega;\mathrm{O}(d)) β ∈ SBV 0 ( Ω ; O ( d )) satisfying (1.16), it holds
F λ ( β , 1 ) ≤ lim inf ε → 0 F ε ( β ε , v ε ) ; \mathcal{F}^{\lambda}(\beta,1)\leq\liminf_{\varepsilon\to0}\mathcal{F}_{\varepsilon}(\beta_{\varepsilon},v_{\varepsilon}); F λ ( β , 1 ) ≤ ε → 0 lim inf F ε ( β ε , v ε ) ; (UB) For any β ∈ S B V 0 ( Ω ; O ( d ) ) \beta\in\mathrm{SBV}_{0}(\Omega;\mathrm{O}(d)) β ∈ SBV 0 ( Ω ; O ( d )) there exists ( β ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d × [ 0 , 1 ] ) (\beta_{\varepsilon},v_{\varepsilon})_{\varepsilon}\subset L^{1}(\Omega;\mathbb{M}^{d\times d}\times[0,1]) ( β ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d × [ 0 , 1 ]) satisfying (1.16) and
F λ ( β , 1 ) ≥ lim sup ε → 0 F ε ( β ε , v ε ) . (17) \mathcal{F}^{\lambda}(\beta,1)\geq\limsup_{\varepsilon\to0}\mathcal{F}_{\varepsilon}(\beta_{\varepsilon},v_{\varepsilon}). \tag*{(17)} F λ ( β , 1 ) ≥ ε → 0 lim sup F ε ( β ε , v ε ) . (17)
Application to image segmentation
To apply Theorem 1.1 to the segmentation of images of polycrystals, we assume N = d N=d N = d and introduce a fidelity term inspired by [8 , 26 ] . Let w ∈ C ( Ω ; R ) w\in C(\Omega;\mathbb{R}) w ∈ C ( Ω ; R ) represent the grayscale image of a polycrystal, and let σ > 0 \sigma>0 σ > 0 be a small parameter of the order of the lattice spacing. We fix a finite number of lattice vectors v 1 , … , v K ∈ R d v_{1},\ldots,v_{K}\in\mathbb{R}^{d} v 1 , … , v K ∈ R d , with ∣ v k ∣ ≤ σ \lvert v_{k}\rvert\leq\sigma ∣ v k ∣ ≤ σ , and positive weights α k > 0 \alpha_{k}>0 α k > 0 , k = 1 , … , K k=1,\ldots,K k = 1 , … , K .
For a given domain Ω ′ \Omega' Ω ′ such that { x : dist ( x , Ω ′ ) < 2 σ } ⋐ Ω \{x:\operatorname{dist}(x,\Omega')<2\sigma\}\Subset\Omega { x : dist ( x , Ω ′ ) < 2 σ } ⋐ Ω we define the fidelity term by
F T ( β ; Ω ′ ) : = ∫ E β ∩ Ω ′ ∑ k = 1 K α k ∑ G ∈ G ( w ( x + ( β ( x ) ) − 1 G v k ) − w ( x ) ) 2 d x , (18) \mathrm{FT}(\beta;\Omega'):=\int_{E_{\beta}\cap\Omega'}\sum_{k=1}^{K}\alpha_{k}\sum_{G\in\mathcal{G}}\left(w\left(x+(\beta(x))^{-1}Gv_{k}\right)-w(x)\right)^{2}\,dx, \tag*{(18)} FT ( β ; Ω ′ ) := ∫ E β ∩ Ω ′ k = 1 ∑ K α k G ∈ G ∑ ( w ( x + ( β ( x ) ) − 1 G v k ) − w ( x ) ) 2 d x , (18)
where
E β : = { x ∈ Ω : det ( β ( x ) ) ≠ 0 , ∥ ( β ( x ) ) − 1 ∥ o p ≤ 2 } , E_{\beta}:=\{x\in\Omega:\det(\beta(x))\neq0,\ \lVert(\beta(x))^{-1}\rVert_{\mathrm{op}}\leq2\}, E β := { x ∈ Ω : det ( β ( x )) = 0 , ∥( β ( x ) ) − 1 ∥ op ≤ 2 } ,
and ∥ ⋅ ∥ o p \lVert\cdot\rVert_{\mathrm{op}} ∥ ⋅ ∥ op denotes the operator norm (i.e., the largest singular value). This definition ensures that the points x x x included in the integral satisfy x + ( β ( x ) ) − 1 G v k ∈ Ω x+(\beta(x))^{-1}Gv_{k}\in\Omega x + ( β ( x ) ) − 1 G v k ∈ Ω , so that F T \mathrm{FT} FT is well-defined.
Since the vectors v k v_k v k belong to the lattice and G \mathcal{G} G is the associated point group, the vectors G v k \mathcal{G}v_k G v k also lie in the lattice. Thus, if w w w is an image of a lattice with orientation R ∈ O ( d ) R \in O(d) R ∈ O ( d ) , then w ( x + R G v k ) = w ( x ) w(x+R\mathcal{G}v_k)=w(x) w ( x + R G v k ) = w ( x ) for all G \mathcal{G} G and k k k , and the fidelity term F T \mathrm{FT} FT is minimized for β = R T \beta=R^{T} β = R T .
The definition of the fidelity term F T \mathrm{FT} FT ensures invariance under the action of the group G \mathcal{G} G . In particular, for every β ∈ S B V G ( Ω ; M d × d ) \beta\in\mathrm{SBV}_{\mathcal{G}}(\Omega;\mathbb{M}^{d\times d}) β ∈ SBV G ( Ω ; M d × d ) and every G ∈ L ∞ ( Ω ; G ) \mathcal{G}\in L^{\infty}(\Omega;\mathcal{G}) G ∈ L ∞ ( Ω ; G ) it holds that
F T ( G β ; Ω ′ ) = F T ( β ; Ω ′ ) . (19) \mathrm{FT}(\mathcal{G}\beta;\Omega')=\mathrm{FT}(\beta;\Omega'). \tag*{(19)} FT ( G β ; Ω ′ ) = FT ( β ; Ω ′ ) . (19)
Moreover, by the continuity of w w w and the fact that any sequence with bounded energy converges to a limit valued in O ( d ) O(d) O ( d ) , which implies ∣ Ω ∖ E β ε ∣ → 0 |\Omega\setminus E_{\beta_\varepsilon}|\to0 ∣Ω ∖ E β ε ∣ → 0 , the fidelity term F T \mathrm{FT} FT is continuous with respect to the L G 1 L_{\mathcal{G}}^{1} L G 1 convergence of β \beta β along sequences with bounded energy. Hence, by Theorem 1.1 , for any fixed γ > 0 \gamma>0 γ > 0 the sequence of functionals
H ε ( β , v ) : = F ε ( β , v ; Ω ) + γ F T ( β ; Ω ′ ) (20) \mathcal{H}_{\varepsilon}(\beta,v):=\mathcal{F}_{\varepsilon}(\beta,v;\Omega)+\gamma\mathrm{FT}(\beta;\Omega') \tag*{(20)} H ε ( β , v ) := F ε ( β , v ; Ω ) + γ FT ( β ; Ω ′ ) (20)
Γ \Gamma Γ -converges (with respect to the L G 1 × L 1 L_{\mathcal{G}}^{1}\times L^{1} L G 1 × L 1 convergence of ( β , v ) (\beta,v) ( β , v ) ) to
H λ ( β , v ) : = F λ ( β , v ; Ω ) + γ F T ( β ; Ω ′ ) . (21) \mathcal{H}^{\lambda}(\beta,v):=\mathcal{F}^{\lambda}(\beta,v;\Omega)+\gamma\mathrm{FT}(\beta;\Omega'). \tag*{(21)} H λ ( β , v ) := F λ ( β , v ; Ω ) + γ FT ( β ; Ω ′ ) . (21)
In view of the equi-coerciveness established in Theorem 1.1 , this also implies convergence of minima and almost minimisers of H ε \mathcal{H}_{\varepsilon} H ε to those of H λ \mathcal{H}^{\lambda} H λ . Clearly the same would also apply to other types of fidelity terms, as far as they have the appropriate invariance under the action of the group G \mathcal{G} G and they are continuous with respect to the L G 1 L_{\mathcal{G}}^{1} L G 1 convergence along sequences with bounded energy.
1.3. Main difficulties and novelties. The key novelty of this paper lies in the presence of the point group G \mathcal{G} G , under whose action the sharp interface model remains invariant. Crucially, we are able to incorporate this invariance also at the level of the approximating functionals. This requires working with an ad hoc notion of convergence for the fields that respects the group action, in appropriate spaces.
A fundamental preliminary challenge is the choice of a suitable functional space for the order parameters. Two main approaches can be considered, both in the formulation of the statement and in the proofs. In the presentation above, we opted for the space S B V G \mathrm{SBV}_{\mathcal{G}} SBV G defined in (1.4), which is a subset of the classical space S B V \mathrm{SBV} SBV . A version of this space (under different notation) appears naturally also in the modeling of fractional dislocations in two dimensions [5 , 31 ] .
An alternative approach is to define the functionals directly on functions taking values in the quotient space M d × d / G \mathbb{M}^{d\times d}/\mathcal{G} M d × d / G . In this framework, one can use the metric structure of M d × d / G \mathbb{M}^{d\times d}/\mathcal{G} M d × d / G and work in the more abstract framework of Sobolev and BV spaces of functions with values in metric spaces, as introduced by Ambrosio in [1 ] (see also [36 ] ). Remarkably, in this metric space approach, the approximating functionals are defined on Sobolev functions as usual in phase field approximation, in particular the fields belong to the space H 1 ( Ω ; M d × d / G ) H^{1}(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}) H 1 ( Ω ; M d × d / G ) . However, this more abstract setting comes with significant difficulties. The lack of linear structure in the quotient space restricts the applicability of many standard analytical and numerical tools, such as the Finite Elements Method. This poses substantial obstacles for both theoretical analysis and numerical simulations.
Our chosen approach thus balances between capturing the essential group invariance and maintaining a workable structure for analysis and approximation.
In order to compare the two approaches, it is important to clarify the relation between S B V G ( Ω ; M d × d ) \mathrm{SBV}_{\mathcal{G}}(\Omega;\mathbb{M}^{d\times d}) SBV G ( Ω ; M d × d ) and H 1 ( Ω ; M d × d / G ) H^{1}(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}) H 1 ( Ω ; M d × d / G ) . It is plain to see that for any function in S B V G ( Ω ; M d × d ) \mathrm{SBV}_{\mathcal{G}}(\Omega;\mathbb{M}^{d\times d}) SBV G ( Ω ; M d × d ) , by passing pointwise to equivalence classes, one obtains a function in H 1 ( Ω ; M d × d / G ) H^{1}(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}) H 1 ( Ω ; M d × d / G ) (cf. Remark 2.4). Conversely, it is not true in dimension d ≥ 2 d\ge2 d ≥ 2 that any function in H 1 ( Ω ; M d × d / G ) H^{1}(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}) H 1 ( Ω ; M d × d / G ) has a (pointwise) representative in S B V G ( Ω ; M d × d ) \mathrm{SBV}_{\mathcal{G}}(\Omega;\mathbb{M}^{d\times d}) SBV G ( Ω ; M d × d ) , since in general the reconstructed jump set may have infinite measure, as in the counterexample of Lemma 2.6 . Therefore, in general, the approximating functionals defined in H 1 ( Ω ; M d × d / G ) H^1(\Omega; \mathrm{M}^{d \times d}/\mathcal{G}) H 1 ( Ω ; M d × d / G ) are smaller than the ones defined in S B V G ( Ω ; M d × d ) \mathrm{SBV}_{\mathcal{G}}(\Omega; \mathrm{M}^{d \times d}) SBV G ( Ω ; M d × d ) , cf. (2.31).
When the reference domain Ω \Omega Ω has dimension 1, instead, its total ordering allows to prove that the two spaces are equivalent, in the sense described just above, and suitably represented by H 1 H^1 H 1 functions with values in M d × d \mathrm{M}^{d \times d} M d × d , see Lemma 2.5. Further, a similar property holds if the target set of the functions is restricted, in particular for the domains of the limiting functionals, which are the spaces S B V G ( Ω ; O ( d ) ) \mathrm{SBV}_{\mathcal{G}}(\Omega; \mathrm{O}(d)) SBV G ( Ω ; O ( d )) and H 1 ( Ω ; O ( d ) / G ) H^1(\Omega; \mathrm{O}(d)/\mathcal{G}) H 1 ( Ω ; O ( d ) / G ) ; in fact, in this case one may work in the framework of Sobolev or BV functions valued into manifolds, and a lifting is possible, see Theorem 2.12 below.
Basing on the former relations, we develop the two approaches in such a way that each one can be followed independently, in particular Theorem 1.1 can be proven without using the theory of Sobolev or BV functions valued in metric spaces, as we now describe.
As for Compactness (C), we employ the convergence of the fields β n \beta_n β n with equibounded energy to a function with values in O ( d ) \mathrm{O}(d) O ( d ) together with the fact that the functions det β n \det\beta_n det β n are equibounded in W 1 , 1 W^{1,1} W 1 , 1 in { v n ≥ c } \{v_n \ge c\} { v n ≥ c } for c > 0 c > 0 c > 0 . The latter follows from the G \mathcal{G} G -invariance of the determinant and the uniform integral bound for ∇ β n \nabla\beta_n ∇ β n in { v n ≥ c } \{v_n \ge c\} { v n ≥ c } . Then, by the lifting result in Theorem 2.12, we obtain a uniform bound in BV, up to modifications in sets with bounded perimeter.
Both in the characterization of the limit domain S B V 0 \mathrm{SBV}_0 SBV 0 and in the Lower Bound (LB) one may reduce to one dimensional slices so, by Lemma 2.5, work with standard Sobolev function and adapt the analogous properties from [19 ] . In the characterization of the limit domain we use here a strategy similar to the one in [21 ] . The Upper Bound (UB) employs a density result in S B V 0 \mathrm{SBV}_0 SBV 0 through functions with jump set given by a finite union of pieces of hyperplanes [7 ] , allowing to use, where the jump is flat, optimal profiles from the one dimensional cell formula for g λ g_\lambda g λ , while a careful construction applies to deal with neighborhoods of the N − 2 N-2 N − 2 dimensional intersections of different faces.
Taking advantage of the abstract theory in [1 ] , one may prove the more general Compactness (C)’ in Theorem 2.15 (recall (2.31)) arguing analogously to the “classical” BV case in [19 ] , while the more general Lower Bound (LB)’ follows similarly to (LB), due to the 1d equivalence from Lemma 2.5 and some fine properties from [1 ] . Conversely, the strongest Upper Bound is (UB), so (UB)’ follows.
The paper is structured as follows. In Section 2 we introduce some notation and recall the basic theory of Sobolev and BV functions valued in metric spaces, discuss the relation with the counterparts which are subspaces of S B V G \mathrm{SBV}_{\mathcal{G}} SBV G , and introduce the formulation for functions valued in equivalence classes along with the main result Theorem 2.15. Then we prove in Section 3 compactness and the characterization of the domain for limiting fields. Section 4 is devoted to define, describe, and characterize the surface tension. Such analysis will be crucial to prove the lower bound and the upper bound, in Section 5.
Formulation with Sobolev functions valued in metric spaces
This Section addresses the formulation of the problem with Sobolev functions valued in metric spaces. We start with some technical preliminaries and notation, which are relevant for the entire paper.
Notation
2.1. Notation. In the entire paper, Ω \Omega Ω is a bounded connected open subset of R N \mathbb{R}^N R N , N ≥ 1 N \ge1 N ≥ 1 , with Lipschitz boundary, and A ( Ω ) \mathcal{A}(\Omega) A ( Ω ) and B ( Ω ) \mathcal{B}(\Omega) B ( Ω ) denote the set of open and Borel subsets of Ω \Omega Ω , respectively. When Ω \Omega Ω is one dimensional we often denote it with I I I . The Lebesgue measure and the k k k -dimensional Hausdorff measure in R N \mathbb{R}^N R N are denoted by L N \mathcal{L}^N L N and H k \mathcal{H}^k H k , respectively. We denote by M d × d \mathrm{M}^{d \times d} M d × d the space of d × d d \times d d × d matrices, with d ≥ 2 d \ge2 d ≥ 2 . For β ∈ M d × d \beta\in\mathrm{M}^{d \times d} β ∈ M d × d , β T \beta^T β T and ∣ β ∣ |\beta| ∣ β ∣ are respectively its transpose matrix and its Euclidean norm.
Given a finite subgroup G G G of O ( d ) O(d) O ( d ) , that will be fixed along the paper, we denote the quotient space of M d × d \mathbb{M}^{d\times d} M d × d with respect to G G G by M d × d / G \mathbb{M}^{d\times d}/G M d × d / G and its elements by
[ β ] = G β : = { G β : G ∈ G } , β ∈ M d × d . [\beta]=G\beta:=\{G\beta:G\in G\},\qquad\beta\in\mathbb{M}^{d\times d}. [ β ] = Gβ := { Gβ : G ∈ G } , β ∈ M d × d .
The space M d × d / G \mathbb{M}^{d\times d}/G M d × d / G is a separable complete metric space endowed with the metric
d G ( [ β ] , [ β ~ ] ) : = min G ∈ G ∣ β − G β ~ ∣ for every [ β ] , [ β ~ ] ∈ M d × d / G . (22) d_G([\beta],[\widetilde{\beta}]):=\min_{G\in G}|\beta-G\widetilde{\beta}|\qquad\text{for every }[\beta],[\widetilde{\beta}]\in\mathbb{M}^{d\times d}/G. \tag*{(22)} d G ([ β ] , [ β ]) := G ∈ G min ∣ β − G β ∣ for every [ β ] , [ β ] ∈ M d × d / G . (22)
Functional setting I: Metric-space valued Sobolev functions
In what follows we always assume that
( X , d X ) is a separable locally compact metric space, (X,d_X)\text{ is a separable locally compact metric space,} ( X , d X ) is a separable locally compact metric space,
having in mind the application to X = M d × d / G X=\mathbb{M}^{d\times d}/G X = M d × d / G or X = O ( d ) / G X=O(d)/G X = O ( d ) / G . The latter is here seen as a subset of the first, endowed with the same metric, as defined in (22) . We observe that O ( d ) / G O(d)/G O ( d ) / G could also be seen as a manifold, then it is natural to adopt on each tangent space the metric that O ( d ) O(d) O ( d ) inherits from M d × d \mathbb{M}^{d\times d} M d × d , or O ( d ) / G O(d)/G O ( d ) / G from M d × d / G \mathbb{M}^{d\times d}/G M d × d / G , and then extend it to each connected component using geodesics; the result is a metric different from, but equivalent to, the one in (22) . The property of X X X being locally compact will be needed for the definition of the space B V ( Ω ; X ) \mathrm{BV}(\Omega;X) BV ( Ω ; X ) , and not for Sobolev spaces.
For every p ∈ [ 1 , + ∞ ] p\in[1,+\infty] p ∈ [ 1 , + ∞ ] , we define the space L p ( Ω ; X ) L^p(\Omega;X) L p ( Ω ; X ) as the space of all Borel functions u : Ω → X u:\Omega\to X u : Ω → X such that there exists a z 0 ∈ X z_0\in X z 0 ∈ X for which d X ( u ( ⋅ ) , z 0 ) : Ω → R d_X(u(\cdot),z_0):\Omega\to\mathbb{R} d X ( u ( ⋅ ) , z 0 ) : Ω → R is in L p ( Ω ) L^p(\Omega) L p ( Ω ) .
Remark 2.1. The L G 1 L^1_G L G 1 -convergence corresponds to the convergence of the equivalence classes in L 1 ( Ω ; M d × d / G ) L^1(\Omega;\mathbb{M}^{d\times d}/G) L 1 ( Ω ; M d × d / G ) , namely
β ε → β in L G 1 ( Ω ; M d × d ) if and only if [ β ε ] → [ β ] in L 1 ( Ω ; M d × d / G ) , \beta_\varepsilon\to\beta\text{ in }L^1_G(\Omega;\mathbb{M}^{d\times d})\quad\text{if and only if}\quad[\beta_\varepsilon]\to[\beta]\text{ in }L^1(\Omega;\mathbb{M}^{d\times d}/G), β ε → β in L G 1 ( Ω ; M d × d ) if and only if [ β ε ] → [ β ] in L 1 ( Ω ; M d × d / G ) ,
which is also equivalent to
lim ε → 0 ∫ Ω d G ( [ β ε ( x ) ] , [ β ( x ) ] ) d x = 0. \lim_{\varepsilon\to0}\int_\Omega d_G([\beta_\varepsilon(x)],[\beta(x)])\,dx=0. ε → 0 lim ∫ Ω d G ([ β ε ( x )] , [ β ( x )]) d x = 0.
Let us consider a function β ∈ S B V G \beta\in\mathrm{SBV}_G β ∈ SBV G , as defined in (4) . Then ∇ β ∈ L 2 \nabla\beta\in L^2 ∇ β ∈ L 2 , and β + ∈ G β − \beta^+\in G\beta^- β + ∈ G β − at almost every jump point. Therefore if we consider the equivalence classes corresponding to β \beta β , defined by [ β ] : Ω → M d × d / G [\beta]:\Omega\to\mathbb{M}^{d\times d}/G [ β ] : Ω → M d × d / G , x ↦ [ β ( x ) ] x\mapsto[\beta(x)] x ↦ [ β ( x )] , the jumps disappear. In particular, this function is expected to be in H 1 ( Ω ; M d × d / G ) H^1(\Omega;\mathbb{M}^{d\times d}/G) H 1 ( Ω ; M d × d / G ) (which has not been defined yet).
To formalize this setting, we recall the notion of Sobolev functions with values in a metric space from [36 ] . By Lip 1 ( X ; R m ) \operatorname{Lip}_1(X;\mathbb{R}^m) Lip 1 ( X ; R m ) we denote the set of Lipschitz functions from X X X to R m \mathbb{R}^m R m with Lipschitz constant Lip ( φ ) \operatorname{Lip}(\varphi) Lip ( φ ) not larger than 1 1 1 , for m = 1 m=1 m = 1 we write Lip 1 ( X ) \operatorname{Lip}_1(X) Lip 1 ( X ) .
Definition 2.2. The space W 1 , p ( Ω ; X ) W^{1,p}(\Omega;X) W 1 , p ( Ω ; X ) , p ≥ 1 p\ge1 p ≥ 1 , consists of those functions u ∈ L p ( Ω ; X ) u\in L^p(\Omega;X) u ∈ L p ( Ω ; X ) such that φ ∘ u ∈ W 1 , p ( Ω ) \varphi\circ u\in W^{1,p}(\Omega) φ ∘ u ∈ W 1 , p ( Ω ) for every φ ∈ Lip ( X ) \varphi\in\operatorname{Lip}(X) φ ∈ Lip ( X ) . For m ∈ N > 0 m\in\mathbb{N}_{>0} m ∈ N > 0 , if we set
∣ ∇ u ∣ ( m ) : = ess sup { ∣ ∇ ( φ ∘ u ) ∣ : φ ∈ Lip 1 ( X ; R m ) } , (23) |\nabla u|_{(m)}:=\operatorname*{ess\,sup}\{|\nabla(\varphi\circ u)|:\varphi\in\operatorname{Lip}_1(X;\mathbb{R}^m)\}, \tag*{(23)} ∣∇ u ∣ ( m ) := ess sup { ∣∇ ( φ ∘ u ) ∣ : φ ∈ Lip 1 ( X ; R m )} , (23)
it holds ∣ ∇ u ∣ ( m ) ∈ L p ( Ω ) |\nabla u|_{(m)}\in L^p(\Omega) ∣∇ u ∣ ( m ) ∈ L p ( Ω ) . We shall mostly take m = d 2 m=d^2 m = d 2 , and when the meaning is clear from the context write briefly ∣ ∇ u ∣ : = ∣ ∇ u ∣ ( d 2 ) |\nabla u|:=|\nabla u|_{(d^2)} ∣∇ u ∣ := ∣∇ u ∣ ( d 2 ) . In dimension N = 1 N=1 N = 1 we use the notation ∣ u ′ ∣ |u'| ∣ u ′ ∣ for ∣ ∇ u ∣ ( 1 ) |\nabla u|_{(1)} ∣∇ u ∣ ( 1 ) .
Remark 2.3. We observe that for any m ∈ N > 0 m\in\mathbb{N}_{>0} m ∈ N > 0
∣ ∇ u ∣ ( 1 ) ≤ ∣ ∇ u ∣ ( m ) ≤ N ∧ m ∣ ∇ u ∣ ( 1 ) pointwise almost everywhere . (24) |\nabla u|_{(1)}\le|\nabla u|_{(m)}\le\sqrt{N\wedge m}\,|\nabla u|_{(1)}\qquad\text{pointwise almost everywhere}. \tag*{(24)} ∣∇ u ∣ ( 1 ) ≤ ∣∇ u ∣ ( m ) ≤ N ∧ m ∣∇ u ∣ ( 1 ) pointwise almost everywhere . (24)
To prove the first inequality we extend any function in Lip 1 ( X ) \operatorname{Lip}_1(X) Lip 1 ( X ) to a function in Lip 1 ( X ; R m ) \operatorname{Lip}_1(X;\mathbb{R}^m) Lip 1 ( X ; R m ) adding m − 1 m-1 m − 1 components which are identically zero. To prove the m \sqrt{m} m bound in the second one, we observe that, if φ = ( φ 1 , … , φ m ) ∈ Lip 1 ( X ; R m ) \varphi=(\varphi_1,\ldots,\varphi_m)\in\operatorname{Lip}_1(X;\mathbb{R}^m) φ = ( φ 1 , … , φ m ) ∈ Lip 1 ( X ; R m ) , then for each i ∈ { 1 , … , m } i\in\{1,\ldots,m\} i ∈ { 1 , … , m } we have φ i ∈ Lip 1 ( X ) \varphi_i\in\operatorname{Lip}_1(X) φ i ∈ Lip 1 ( X ) , and ∣ ∇ ( φ ∘ u ) ∣ = ∣ ( ∇ ( φ 1 ∘ u ) , … , ∇ ( φ m ∘ u ) ) ∣ ≤ m max i ∣ ∇ ( φ i ∘ u ) ∣ |\nabla(\varphi\circ u)|=|(\nabla(\varphi_1\circ u),\ldots,\nabla(\varphi_m\circ u))|\le\sqrt{m}\max_i|\nabla(\varphi_i\circ u)| ∣∇ ( φ ∘ u ) ∣ = ∣ ( ∇ ( φ 1 ∘ u ) , … , ∇ ( φ m ∘ u )) ∣ ≤ m max i ∣∇ ( φ i ∘ u ) ∣ . To prove
verts the page end. the final inequality, let φ ∈ Lip 1 ( X ; R m ) \varphi\in\operatorname{Lip}_{1}(X;\mathbb{R}^{m}) φ ∈ Lip 1 ( X ; R m ) , fix v ∈ S N − 1 v \in S^{N-1} v ∈ S N − 1 , and observe that for almost every x ∈ Ω x \in\Omega x ∈ Ω and t t t small enough, writing I t I_t I t the segment with endpoints 0 0 0 and t t t , one has
∣ φ ( u ( x + t v ) ) − φ ( u ( x ) ) ∣ ≤ d X ( u ( x + t v ) , u ( x ) ) ≤ ∫ I t ∣ ∇ ( d X ∘ u ) ∣ ( x + s v ) d s ≤ ∫ I t ∣ ∇ u ∣ ( 1 ) ( x + s v ) d s . (25) \begin{aligned}
|\varphi(u(x+tv))-\varphi(u(x))| \le d_X(u(x+tv),u(x)) \le\int_{I_t} |\nabla(d_X\circ u)|(x+sv)\,ds \\
&\le\int_{I_t} |\nabla u|_{(1)}(x+sv)\,ds. \tag*{(25)}
\end{aligned} ∣ φ ( u ( x + t v )) − φ ( u ( x )) ∣ ≤ d X ( u ( x + t v ) , u ( x )) ≤ ∫ I t ∣∇ ( d X ∘ u ) ∣ ( x + s v ) d s ≤ ∫ I t ∣∇ u ∣ ( 1 ) ( x + s v ) d s . (25)
Therefore if x x x is a Lebesgue point of the slice of ∣ ∇ u ∣ ( 1 ) |\nabla u|_{(1)} ∣∇ u ∣ ( 1 ) , then
lim sup t → 0 ∣ φ ( u ( x + t v ) ) − φ ( u ( x ) ) ∣ ∣ t ∣ ≤ lim sup t → 0 1 ∣ t ∣ ∫ I t ∣ ∇ u ∣ ( 1 ) ( x + s v ) d s = ∣ ∇ u ∣ ( 1 ) ( x ) (26) \limsup_{t\to0}\frac{|\varphi(u(x+tv))-\varphi(u(x))|}{|t|}
\le\limsup_{t\to0}\frac{1}{|t|}\int_{I_t}|\nabla u|_{(1)}(x+sv)\,ds
=|\nabla u|_{(1)}(x) \tag*{(26)} t → 0 lim sup ∣ t ∣ ∣ φ ( u ( x + t v )) − φ ( u ( x )) ∣ ≤ t → 0 lim sup ∣ t ∣ 1 ∫ I t ∣∇ u ∣ ( 1 ) ( x + s v ) d s = ∣∇ u ∣ ( 1 ) ( x ) (26)
and we obtain ∣ ∇ v ( φ ∘ u ) ∣ ≤ ∣ ∇ u ∣ ( 1 ) |\nabla_v(\varphi\circ u)|\le|\nabla u|_{(1)} ∣ ∇ v ( φ ∘ u ) ∣ ≤ ∣∇ u ∣ ( 1 ) almost everywhere. Finally, for any g ∈ W 1 , p ( Ω ; R m ) g\in W^{1,p}(\Omega;\mathbb{R}^{m}) g ∈ W 1 , p ( Ω ; R m ) we have
∣ ∇ g ∣ 2 ( x ) = ∑ i ∣ ∂ i g ∣ 2 ( x ) ≤ N max i ∣ ∂ i g ∣ 2 ( x ) . (27) |\nabla g|^{2}(x)=\sum_i|\partial_i g|^{2}(x)\le N\max_i|\partial_i g|^{2}(x). \tag*{(27)} ∣∇ g ∣ 2 ( x ) = i ∑ ∣ ∂ i g ∣ 2 ( x ) ≤ N i max ∣ ∂ i g ∣ 2 ( x ) . (27)
Using this with g = φ ∘ u g=\varphi\circ u g = φ ∘ u , we obtain ∣ ∇ ( φ ∘ u ) ∣ ( m ) ≤ N ∣ ∇ u ∣ ( 1 ) |\nabla(\varphi\circ u)|_{(m)}\le\sqrt{N}|\nabla u|_{(1)} ∣∇ ( φ ∘ u ) ∣ ( m ) ≤ N ∣∇ u ∣ ( 1 ) almost everywhere which concludes the proof of (2.3).
We next show that
if X ⊂ R m , then ∣ ∇ u ∣ ( m ) = ∣ ∇ u ∣ R m × N , (28) \text{if }X\subset\mathbb{R}^{m},\text{ then }|\nabla u|_{(m)}=|\nabla u|_{\mathbb{R}^{m\times N}}, \tag*{(28)} if X ⊂ R m , then ∣∇ u ∣ ( m ) = ∣∇ u ∣ R m × N , (28)
where the last norm is the usual Euclidean norm of the weak gradient. To see this, observe that any φ ∈ Lip 1 ( X ; R m ) \varphi\in\operatorname{Lip}_{1}(X;\mathbb{R}^{m}) φ ∈ Lip 1 ( X ; R m ) has an extension φ ~ ∈ Lip 1 ( R m ; R m ) \widetilde{\varphi}\in\operatorname{Lip}_{1}(\mathbb{R}^{m};\mathbb{R}^{m}) φ ∈ Lip 1 ( R m ; R m ) , and by the chain rule ∣ ∇ ( φ ~ ∘ u ) ∣ = ∣ ( ∇ φ ~ ) ∘ u ∇ u ∣ ≤ ∣ ∇ u ∣ |\nabla(\widetilde{\varphi}\circ u)|=|(\nabla\widetilde{\varphi})\circ u\,\nabla u|\le|\nabla u| ∣∇ ( φ ∘ u ) ∣ = ∣ ( ∇ φ ) ∘ u ∇ u ∣ ≤ ∣∇ u ∣ . At the same time, for φ ( x ) = x \varphi(x)=x φ ( x ) = x the inequality is an equality, which proves (2.7). A similar argument shows that the bounds in (2.3) are optimal.
Remark 2.4. The functions in SBV G ( Ω ; M d × d ) \operatorname{SBV}_{\mathcal{G}}(\Omega;M^{d\times d}) SBV G ( Ω ; M d × d ) are strictly related to the functions in the Sobolev space H 1 ( Ω ; M d × d / G ) : = W 1 , 2 ( Ω ; M d × d / G ) H^{1}(\Omega;M^{d\times d}/G):=W^{1,2}(\Omega;M^{d\times d}/G) H 1 ( Ω ; M d × d / G ) := W 1 , 2 ( Ω ; M d × d / G ) . To see this, for a Borel function β : Ω → M d × d \beta:\Omega\to M^{d\times d} β : Ω → M d × d we define
[ β ] : Ω → M d × d / G , [ β ] ( x ) : = [ β ( x ) ] , x ∈ Ω . [\beta]:\Omega\to M^{d\times d}/G,\qquad[\beta](x):=[\beta(x)],\quad x\in\Omega. [ β ] : Ω → M d × d / G , [ β ] ( x ) := [ β ( x )] , x ∈ Ω.
By the chain rule in BV, if β ∈ SBV G ( Ω ; M d × d ) \beta\in\operatorname{SBV}_{\mathcal{G}}(\Omega;M^{d\times d}) β ∈ SBV G ( Ω ; M d × d ) and φ ∈ Lip ( M d × d / G ; R m ) \varphi\in\operatorname{Lip}(M^{d\times d}/G;\mathbb{R}^{m}) φ ∈ Lip ( M d × d / G ; R m ) , then φ ∘ [ β ] \varphi\circ[\beta] φ ∘ [ β ] belongs to H 1 ( Ω ; R m ) H^{1}(\Omega;\mathbb{R}^{m}) H 1 ( Ω ; R m ) , and
∣ ∇ ( φ ∘ [ β ] ) ∣ ≤ Lip ( φ ) ∣ ∇ β ∣ , a.e. in Ω , (29) |\nabla(\varphi\circ[\beta])|\le\operatorname{Lip}(\varphi)|\nabla\beta|,\qquad\text{a.e. in }\Omega, \tag*{(29)} ∣∇ ( φ ∘ [ β ]) ∣ ≤ Lip ( φ ) ∣∇ β ∣ , a.e. in Ω , (29)
so that [ β ] ∈ H 1 ( Ω ; M d × d / G ) [\beta]\in H^{1}(\Omega;M^{d\times d}/G) [ β ] ∈ H 1 ( Ω ; M d × d / G ) , and from (2.8)
∣ ∇ [ β ] ∣ ( m ) ≤ ∣ ∇ β ∣ , m ∈ N > 0 . (30) |\nabla[\beta]|_{(m)}\le|\nabla\beta|,\qquad m\in\mathbb{N}_{>0}. \tag*{(30)} ∣∇ [ β ] ∣ ( m ) ≤ ∣∇ β ∣ , m ∈ N > 0 . (30)
Hence the functions in SBV G \operatorname{SBV}_{\mathcal{G}} SBV G can be identified with Sobolev functions with values in the quotient space M d × d / G M^{d\times d}/G M d × d / G .
If the domain is one dimensional, the two spaces SBV G ( I ; M d × d ) \operatorname{SBV}_{\mathcal{G}}(I;M^{d\times d}) SBV G ( I ; M d × d ) and H 1 ( I ; M d × d / G ) H^{1}(I;M^{d\times d}/G) H 1 ( I ; M d × d / G ) are actually equivalent. We can moreover “clean the jumps” of a SBV G ( I ; M d × d ) \operatorname{SBV}_{\mathcal{G}}(I;M^{d\times d}) SBV G ( I ; M d × d ) function, obtaining a representative in H 1 ( I ; M d × d ) H^{1}(I;M^{d\times d}) H 1 ( I ; M d × d ) , as shown in Lemma 2.5 below. This result will be crucial for our proof.
If the domain is at least two dimensional, the two spaces are not equivalent, see Lemma 2.6 below. However, the counterexample is based on the fact that functions in SBV G ( I ; M d × d ) \operatorname{SBV}_{\mathcal{G}}(I;M^{d\times d}) SBV G ( I ; M d × d ) are required to have jump set of finite measure. If this property is eliminated from the definition, then we expect that it is possible to identify the two spaces (see Remark 2.14 below).
Lemma 2.5. Let I = ( a , b ) ⊂ R I=(a,b)\subset\mathbb{R} I = ( a , b ) ⊂ R be a bounded interval, and let β : I → M d × d \beta:I\to M^{d\times d} β : I → M d × d be a Borel function. Then the following hold:
i) if [ β ] ∈ H 1 ( I ; M d × d / G ) [\beta]\in H^{1}(I;M^{d\times d}/G) [ β ] ∈ H 1 ( I ; M d × d / G ) , then there exists G β ∈ L ∞ ( I ; G ) G_{\beta}\in L^{\infty}(I;G) G β ∈ L ∞ ( I ; G ) such that G β β ∈ H 1 ( I ; M d × d ) G_{\beta}\beta\in H^{1}(I;M^{d\times d}) G β β ∈ H 1 ( I ; M d × d ) , G β β ( a ) = β ( a ) G_{\beta}\beta(a)=\beta(a) G β β ( a ) = β ( a ) , and ∣ [ β ] ′ ∣ ( m ) = ∣ ( G β β ) ′ ∣ |[\beta]'|_{(m)}=|(G_{\beta}\beta)'| ∣ [ β ] ′ ∣ ( m ) = ∣ ( G β β ) ′ ∣ almost everywhere in I I I for any m ∈ N > 0 m\in\mathbb{N}_{>0} m ∈ N > 0 ; ii) if, in addition, β ∈ S B V G ( I ; M d × d ) \beta\in\mathrm{SBVG}(I; \mathbb{M}^{d \times d}) β ∈ SBVG ( I ; M d × d ) , then ( G β β ) ′ = G β β ′ (G_{\beta}\beta)' = G_{\beta}\beta' ( G β β ) ′ = G β β ′ almost everywhere in I I I .
Proof. Let β : I → M d × d \beta: I \to\mathbb{M}^{d \times d} β : I → M d × d be such that [ β ] ( x ) : = [ β ( x ) ] [\beta](x) := [\beta(x)] [ β ] ( x ) := [ β ( x )] belongs to H 1 ( I ; M d × d / G ) H^1(I; \mathbb{M}^{d \times d}/G) H 1 ( I ; M d × d / G ) , and therefore it admits a continuous representative (see, e.g. [1, Remark 2.2]). First of all, we prove that [ β ] ∈ C 1 / 2 ( I ; M d × d / G ) [\beta] \in C^{1/2}(I; \mathbb{M}^{d \times d}/G) [ β ] ∈ C 1/2 ( I ; M d × d / G ) . To see this, fix x ∈ I x \in I x ∈ I , and consider φ x ( U ) : = d G ( U , [ β ] ( x ) ) \varphi_x(U) := d_G(U, [\beta](x)) φ x ( U ) := d G ( U , [ β ] ( x )) , U ∈ M d × d / G U \in\mathbb{M}^{d \times d}/G U ∈ M d × d / G . Since φ x \varphi_x φ x is 1-Lipschitz, and [ β ] ∈ H 1 ( I ; M d × d / G ) [\beta] \in H^1(I; \mathbb{M}^{d \times d}/G) [ β ] ∈ H 1 ( I ; M d × d / G ) , we have
d G ( [ β ] ( x ) , [ β ] ( y ) ) = φ x ( [ β ] ( y ) ) − φ x ( [ β ] ( x ) ) ≤ ∫ [ x , y ] ∣ ( φ x ∘ [ β ] ) ′ ∣ d t ≤ ∣ x − y ∣ 1 / 2 ∥ ( φ x ∘ [ β ] ) ′ ∥ L 2 ( I ) ≤ ∣ x − y ∣ 1 / 2 ∥ ∣ [ β ] ′ ∣ ( m ) ∥ L 2 ( I ) . \begin{aligned}
d_G([\beta](x), [\beta](y)) = \varphi_x([\beta](y)) - \varphi_x([\beta](x)) \le\int_{[x,y]} |(\varphi_x \circ[\beta])'|\,dt \\
&\le|x-y|^{1/2}\|(\varphi_x \circ[\beta])'\|_{L^2(I)} \le|x-y|^{1/2}\| |[\beta]'|_{(m)}\|_{L^2(I)}.
\end{aligned} d G ([ β ] ( x ) , [ β ] ( y )) = φ x ([ β ] ( y )) − φ x ([ β ] ( x )) ≤ ∫ [ x , y ] ∣ ( φ x ∘ [ β ] ) ′ ∣ d t ≤ ∣ x − y ∣ 1/2 ∥ ( φ x ∘ [ β ] ) ′ ∥ L 2 ( I ) ≤ ∣ x − y ∣ 1/2 ∥∣ [ β ] ′ ∣ ( m ) ∥ L 2 ( I ) .
In order to prove i), we first construct a sequence ( β k ) k (\beta_k)_k ( β k ) k of functions in H 1 ( I ; M d × d ) H^1(I; \mathbb{M}^{d \times d}) H 1 ( I ; M d × d ) such that
∥ β k ′ ∥ L 2 ( I ) 2 ≤ ∥ ∣ [ β ] ′ ∣ ( m ) 2 ∥ L 2 ( I ) , β k ( a ) = β ( a ) , ∥ d G ( [ β k ] , [ β ] ) ∥ L ∞ ( I ) → 0. (31) \|\beta_k'\|_{L^2(I)}^2 \le\|\lvert[\beta]'\rvert_{(m)}^2\|_{L^2(I)}, \qquad\beta_k(a) = \beta(a), \qquad\|d_G([\beta_k],[\beta])\|_{L^\infty(I)} \to0. \tag*{(31)} ∥ β k ′ ∥ L 2 ( I ) 2 ≤ ∥ ∣[ β ] ′ ∣ ( m ) 2 ∥ L 2 ( I ) , β k ( a ) = β ( a ) , ∥ d G ([ β k ] , [ β ]) ∥ L ∞ ( I ) → 0. (31)
To do this, we subdivide the interval I I I in k k k subintervals, choosing a 0 : = a < a 1 < ⋯ < a k : = b a_0 := a < a_1 < \cdots< a_k := b a 0 := a < a 1 < ⋯ < a k := b such that ∣ a i + 1 − a i ∣ ≤ 2 ( b − a ) / k |a_{i+1}-a_i| \le2(b-a)/k ∣ a i + 1 − a i ∣ ≤ 2 ( b − a ) / k for all i i i . Further, for i ≥ 1 i \ge1 i ≥ 1 the function φ i : M d × d → R \varphi_i : \mathbb{M}^{d \times d} \to\mathbb{R} φ i : M d × d → R defined by φ i ( U ) : = d G ( [ β ( a i − 1 ) ] , U ) \varphi_i(U) := d_G([\beta(a_{i-1})],U) φ i ( U ) := d G ([ β ( a i − 1 )] , U ) is 1-Lipschitz, hence φ i ∘ [ β ] ∈ H 1 ( I ) \varphi_i \circ[\beta] \in H^1(I) φ i ∘ [ β ] ∈ H 1 ( I ) and, by (2.2),
φ i 2 ( [ β ( a i ) ] ) ≤ ( a i − a i − 1 ) ∫ a i − 1 a i ∣ d d t ( φ i ∘ [ β ] ) ∣ 2 d t ≤ ( a i − a i − 1 ) ∫ a i − 1 a i ∣ [ β ] ′ ∣ ( m ) 2 d t . (32) \varphi_i^2([\beta(a_i)]) \le(a_i-a_{i-1})\int_{a_{i-1}}^{a_i}\left|\frac{\mathrm{d}}{\mathrm{d}t}(\varphi_i \circ[\beta])\right|^2\,\mathrm{d}t \le(a_i-a_{i-1})\int_{a_{i-1}}^{a_i}|[\beta]'|_{(m)}^2\,\mathrm{d}t. \tag*{(32)} φ i 2 ([ β ( a i )]) ≤ ( a i − a i − 1 ) ∫ a i − 1 a i d t d ( φ i ∘ [ β ]) 2 d t ≤ ( a i − a i − 1 ) ∫ a i − 1 a i ∣ [ β ] ′ ∣ ( m ) 2 d t . (32)
We set β k ( a 0 ) = β ( a 0 ) \beta_k(a_0)=\beta(a_0) β k ( a 0 ) = β ( a 0 ) , then inductively select β k ( a i ) ∈ [ β ] ( a i ) \beta_k(a_i) \in[\beta](a_i) β k ( a i ) ∈ [ β ] ( a i ) such that
∣ β k ( a i ) − β k ( a i − 1 ) ∣ = d G ( [ β ] ( a i ) , [ β ] ( a i − 1 ) ) = φ i ( [ β ] ( a i ) ) , (33) |\beta_k(a_i)-\beta_k(a_{i-1})|=d_G([\beta](a_i),[\beta](a_{i-1}))=\varphi_i([\beta](a_i)), \tag*{(33)} ∣ β k ( a i ) − β k ( a i − 1 ) ∣ = d G ([ β ] ( a i ) , [ β ] ( a i − 1 )) = φ i ([ β ] ( a i )) , (33)
and define the function β k \beta_k β k as the affine interpolation in each of the intervals. By (2.12) and (2.11) we get
∫ a i − 1 a i ∣ β k ′ ∣ 2 d t = 1 a i − a i − 1 ∣ β k ( a i ) − β k ( a i − 1 ) ∣ 2 ≤ ∫ a i − 1 a i ∣ [ β ] ′ ∣ ( m ) 2 d t . (34) \int_{a_{i-1}}^{a_i}|\beta_k'|^2\,\mathrm{d}t = \frac{1}{a_i-a_{i-1}}|\beta_k(a_i)-\beta_k(a_{i-1})|^2 \le\int_{a_{i-1}}^{a_i}|[\beta]'|_{(m)}^2\,\mathrm{d}t. \tag*{(34)} ∫ a i − 1 a i ∣ β k ′ ∣ 2 d t = a i − a i − 1 1 ∣ β k ( a i ) − β k ( a i − 1 ) ∣ 2 ≤ ∫ a i − 1 a i ∣ [ β ] ′ ∣ ( m ) 2 d t . (34)
and the first condition in (2.10) follows. The second one holds by the choice of β k ( a 0 ) \beta_k(a_0) β k ( a 0 ) . The third one follows by the embedding of H 1 H^1 H 1 in C 1 / 2 C^{1/2} C 1/2 , with a computation similar to (2.11). More explicitly, we obtain for any t ∈ ( a i − 1 , a i ) t \in(a_{i-1},a_i) t ∈ ( a i − 1 , a i ) ,
d G ( [ β k ] ( t ) , [ β ] ( t ) ) ≤ ∣ β k ( t ) − β k ( a i − 1 ) ∣ + φ i ( [ β ] ( t ) ) ≤ φ i ( [ β ] ( a i ) ) + φ i ( [ β ] ( t ) ) d_G([\beta_k](t),[\beta](t)) \le|\beta_k(t)-\beta_k(a_{i-1})|+\varphi_i([\beta](t)) \le\varphi_i([\beta](a_i))+\varphi_i([\beta](t)) d G ([ β k ] ( t ) , [ β ] ( t )) ≤ ∣ β k ( t ) − β k ( a i − 1 ) ∣ + φ i ([ β ] ( t )) ≤ φ i ([ β ] ( a i )) + φ i ([ β ] ( t ))
≤ 2 a i − a i − 1 ( ∫ a i − 1 a i ∣ [ β ] ′ ∣ ( m ) 2 d t ) 1 / 2 ≤ C k , \le2\sqrt{a_i-a_{i-1}}\left(\int_{a_{i-1}}^{a_i}|[\beta]'|_{(m)}^2\,\mathrm{d}t\right)^{1/2} \le\frac{C}{\sqrt{k}}, ≤ 2 a i − a i − 1 ( ∫ a i − 1 a i ∣ [ β ] ′ ∣ ( m ) 2 d t ) 1/2 ≤ k C ,
and therefore the uniform convergence to zero in (2.10) is proved.
Now by compactness, one obtains that ( β k ) k (\beta_k)_k ( β k ) k has a subsequence converging weakly in H 1 ( I ; M d × d ) H^1(I; \mathbb{M}^{d \times d}) H 1 ( I ; M d × d ) (and almost everywhere in I I I ) to some β ∗ ∈ H 1 ( I ; M d × d ) \beta_* \in H^1(I; \mathbb{M}^{d \times d}) β ∗ ∈ H 1 ( I ; M d × d ) , with β ∗ ( a ) = β ( a ) \beta_*(a)=\beta(a) β ∗ ( a ) = β ( a ) . By the continuity of the distance d G d_G d G we obtain [ β ∗ ] = [ β ] [\beta_*]=[\beta] [ β ∗ ] = [ β ] almost everywhere, and hence there exists G β : I → G G_\beta: I \to G G β : I → G such that β ∗ = G β β \beta_*=G_\beta\beta β ∗ = G β β almost everywhere. Since G G G is finite one immediately obtains that G β G_\beta G β can be chosen measurable. By the lower semicontinuity of the norm, and (2.13) we obtain
∫ I ∣ β ∗ ′ ∣ 2 d t ≤ ∫ I ∣ [ β ] ′ ∣ ( m ) 2 d t = ∫ I ∣ [ β ∗ ] ′ ∣ ( m ) 2 d t \int_I|\beta_*'|^2\,\mathrm{d}t \le\int_I|[\beta]'|_{(m)}^2\,\mathrm{d}t = \int_I|[\beta_*]'|_{(m)}^2\,\mathrm{d}t ∫ I ∣ β ∗ ′ ∣ 2 d t ≤ ∫ I ∣ [ β ] ′ ∣ ( m ) 2 d t = ∫ I ∣ [ β ∗ ] ′ ∣ ( m ) 2 d t
that, together with (2.9) (applied to β ∗ \beta_* β ∗ ), implies that ∣ β ∗ ′ ∣ = ∣ [ β ∗ ] ′ ∣ ( m ) |\beta_*'|=|[\beta_*]'|_{(m)} ∣ β ∗ ′ ∣ = ∣ [ β ∗ ] ′ ∣ ( m ) a.e. in I I I , concluding the proof of the first part of the Lemma.
Now we assume, in addition, that β \beta β belongs to S B V G ( I ; M d × d ) \mathrm{SBV}_{G}(I;M^{d\times d}) SBV G ( I ; M d × d ) , and we show that ( G β β ) ′ = G β β ′ (G_{\beta}\beta)'=G_{\beta}\beta' ( G β β ) ′ = G β β ′ almost everywhere in I I I . To do this, fix any t ∈ I t\in I t ∈ I which is a point of approximate differentiability of β \beta β and β ∗ \beta_{*} β ∗ , so that
lim r → 0 ∫ t − r t + r ∣ β ( s ) − β ( t ) − β ′ ( t ) ( s − t ) ∣ r 2 + ∣ β ∗ ( s ) − β ∗ ( t ) − β ∗ ′ ( t ) ( s − t ) ∣ r 2 d s = 0. (35) \lim_{r\to0}\int_{t-r}^{t+r}\frac{|\beta(s)-\beta(t)-\beta'(t)(s-t)|}{r^{2}}+\frac{|\beta_{*}(s)-\beta_{*}(t)-\beta_{*}'(t)(s-t)|}{r^{2}}\,\mathrm{d}s=0. \tag*{(35)} r → 0 lim ∫ t − r t + r r 2 ∣ β ( s ) − β ( t ) − β ′ ( t ) ( s − t ) ∣ + r 2 ∣ β ∗ ( s ) − β ∗ ( t ) − β ∗ ′ ( t ) ( s − t ) ∣ d s = 0. (35)
If t t t is also a Lebesgue point of G β G_{\beta} G β , as G G G is discrete, we obtain that the set
E r : = { s ∈ ( t − r , t + r ) : G β ( s ) = G β ( t ) } E_{r}:=\{s\in(t-r,t+r):G_{\beta}(s)=G_{\beta}(t)\} E r := { s ∈ ( t − r , t + r ) : G β ( s ) = G β ( t )}
obeys ∣ E r ∣ / r → 2 |E_{r}|/r\to2 ∣ E r ∣/ r → 2 as r → 0 r\to0 r → 0 . Using β ∗ ( s ) = G β ( t ) β ( s ) \beta_{*}(s)=G_{\beta}(t)\beta(s) β ∗ ( s ) = G β ( t ) β ( s ) in E r E_{r} E r and (2.14) we obtain
lim r → 0 ∫ E r ∣ ( G β ( t ) β ′ ( t ) − β ∗ ′ ( t ) ) ( s − t ) ∣ r 2 d s = 0 \lim_{r\to0}\int_{E_{r}}\frac{|(G_{\beta}(t)\beta'(t)-\beta_{*}'(t))(s-t)|}{r^{2}}\,\mathrm{d}s=0 r → 0 lim ∫ E r r 2 ∣ ( G β ( t ) β ′ ( t ) − β ∗ ′ ( t )) ( s − t ) ∣ d s = 0
which implies the desired identity G β ( t ) β ′ ( t ) = β ∗ ′ ( t ) G_{\beta}(t)\beta'(t)=\beta_{*}'(t) G β ( t ) β ′ ( t ) = β ∗ ′ ( t ) .
The next Lemma shows with an example that there are functions in S B V ( Ω ; M 2 × 2 / G ) \mathrm{SBV}(\Omega;M^{2\times2}/G) SBV ( Ω ; M 2 × 2 / G ) , with N , d ≥ 2 N,d\ge2 N , d ≥ 2 , for which we cannot find a representative with jump set of finite measure. Therefore in dimension at least two the result proved in Lemma 2.5 does not hold.
Lemma 2.6. If N ≥ 2 N\ge2 N ≥ 2 , d ≥ 2 d\ge2 d ≥ 2 , and G ∩ S O ( d ) ≠ { I d } G\cap\mathrm{SO}(d)\ne\{\mathrm{Id}\} G ∩ SO ( d ) = { Id } , then there is u ∈ H 1 ( ( 0 , 1 ) N ; M d × d / G ) u\in H^{1}((0,1)^{N};M^{d\times d}/G) u ∈ H 1 (( 0 , 1 ) N ; M d × d / G ) such that there is no β 0 ∈ S B V G ( ( 0 , 1 ) N ; M d × d ) \beta_{0}\in\mathrm{SBV}_{G}((0,1)^{N};M^{d\times d}) β 0 ∈ SBV G (( 0 , 1 ) N ; M d × d ) with u = [ β 0 ] u=[\beta_{0}] u = [ β 0 ] .
Proof. We provide an explicit construction in the case N = 2 N=2 N = 2 , the case N > 2 N>2 N > 2 follows by extending the construction so that it does not depend on the variables x 3 , … , x N x_{3},\ldots,x_{N} x 3 , … , x N , in the sense that u ( N ) ( x ) = u ( 2 ) ( x 1 , x 2 ) u^{(N)}(x)=u^{(2)}(x_{1},x_{2}) u ( N ) ( x ) = u ( 2 ) ( x 1 , x 2 ) .
Select G ∈ G ∩ S O ( d ) G\in G\cap\mathrm{SO}(d) G ∈ G ∩ SO ( d ) , G ≠ I d G\ne\mathrm{Id} G = Id , and then pick a curve γ ∈ C 1 ( [ 0 , 2 π ] ; S O ( d ) ) \gamma\in C^{1}([0,2\pi];\mathrm{SO}(d)) γ ∈ C 1 ([ 0 , 2 π ] ; SO ( d )) such that γ ( 0 ) = I d \gamma(0)=\mathrm{Id} γ ( 0 ) = Id , γ ( 2 π ) = G \gamma(2\pi)=G γ ( 2 π ) = G . This exists since S O ( d ) \mathrm{SO}(d) SO ( d ) is connected. Fix φ ∈ C c 1 ( ( 0 , 1 ) ; [ 0 , 1 ] ) \varphi\in C^{1}_{c}((0,1);[0,1]) φ ∈ C c 1 (( 0 , 1 ) ; [ 0 , 1 ]) such that { φ > 0 } = ( 1 / 4 , 3 / 4 ) \{\varphi>0\}=(1/4,3/4) { φ > 0 } = ( 1/4 , 3/4 ) . We start working in the unit ball B 1 B_{1} B 1 and use polar coordinates. We define, for t ∈ ( 0 , 1 ) t\in(0,1) t ∈ ( 0 , 1 ) and θ ∈ [ 0 , 2 π ) \theta\in[0,2\pi) θ ∈ [ 0 , 2 π ) ,
β 1 ( t e i θ ) : = φ ( t ) γ ( θ ) . \beta_{1}(t e^{i\theta}):=\varphi(t)\gamma(\theta). β 1 ( t e i θ ) := φ ( t ) γ ( θ ) .
We observe that β 1 ∈ S B V ( B 1 ; M d × d ) \beta_{1}\in\mathrm{SBV}(B_{1};M^{d\times d}) β 1 ∈ SBV ( B 1 ; M d × d ) , with β 1 = 0 \beta_{1}=0 β 1 = 0 around ∂ B 1 \partial B_{1} ∂ B 1 , ∇ β 1 ∈ L ∞ ( B 1 ) \nabla\beta_{1}\in L^{\infty}(B_{1}) ∇ β 1 ∈ L ∞ ( B 1 ) , J β 1 = ( 1 / 4 , 3 / 4 ) e 1 J_{\beta_{1}}=(1/4,3/4)e_{1} J β 1 = ( 1/4 , 3/4 ) e 1 , β 1 + ( t e 1 ) = φ ( t ) G \beta_{1}^{+}(te_{1})=\varphi(t)G β 1 + ( t e 1 ) = φ ( t ) G ; β 1 − ( t e 1 ) = φ ( t ) I d \beta_{1}^{-}(te_{1})=\varphi(t)\mathrm{Id} β 1 − ( t e 1 ) = φ ( t ) Id , so that β 1 ∈ S B V G ( B 1 ; M d × d ) \beta_{1}\in\mathrm{SBV}_{G}(B_{1};M^{d\times d}) β 1 ∈ SBV G ( B 1 ; M d × d ) . We define
c 2 : = ∥ ∇ β 1 ∥ L 2 ( B 1 ) 2 ∈ ( 0 , ∞ ) , c J : = H 1 ( J β 1 ∩ B 1 ) = 1 2 , c D : = ∫ 0 1 ∣ β 1 + − β 1 − ∣ ( t e 1 ) d t ∈ ( 0 , ∞ ) . c_{2}:=\|\nabla\beta_{1}\|_{L^{2}(B_{1})}^{2}\in(0,\infty),\qquad c_{J}:=\mathcal{H}^{1}(J_{\beta_{1}}\cap B_{1})=\frac{1}{2},\qquad c_{D}:=\int_{0}^{1}|\beta_{1}^{+}-\beta_{1}^{-}|(te_{1})\,\mathrm{d}t\in(0,\infty). c 2 := ∥∇ β 1 ∥ L 2 ( B 1 ) 2 ∈ ( 0 , ∞ ) , c J := H 1 ( J β 1 ∩ B 1 ) = 2 1 , c D := ∫ 0 1 ∣ β 1 + − β 1 − ∣ ( t e 1 ) d t ∈ ( 0 , ∞ ) .
Consider countably many pairwise disjoint balls B k : = B r k ( x k ) B_{k}:=B_{r_{k}}(x_{k}) B k := B r k ( x k ) contained in the unit square Q : = ( 0 , 1 ) 2 Q:=(0,1)^{2} Q := ( 0 , 1 ) 2 such that
∑ k r k = ∞ . \sum_{k}r_{k}=\infty. k ∑ r k = ∞.
To choose them one can, for example, take any dyadic decomposition of the unit square into countably many subsquares and consider the circle inscribed in each subsquare. We set, for a : N → ( 0 , ∞ ) a:\mathbb{N}\to(0,\infty) a : N → ( 0 , ∞ ) chosen below and K ∈ N K\in\mathbb{N} K ∈ N ,
β K ( x ) : = ∑ k = 1 K a k χ B k ( x ) β 1 ( x − x k r k ) \beta_{K}(x):=\sum_{k=1}^{K}a_{k}\chi_{B_{k}}(x)\beta_{1}\left(\frac{x-x_{k}}{r_{k}}\right) β K ( x ) := k = 1 ∑ K a k χ B k ( x ) β 1 ( r k x − x k )
and verify that β K ∈ S B V G ( Q ; M d × d ) \beta_{K}\in\mathrm{SBV}_{G}(Q;M^{d\times d}) β K ∈ SBV G ( Q ; M d × d ) , with [ β K ] ∈ H 1 ( Q ; M d × d / G ) [\beta_{K}]\in H^{1}(Q;M^{d\times d}/G) [ β K ] ∈ H 1 ( Q ; M d × d / G ) . We compute
∥ ∇ β K ∥ L 2 ( Q ) 2 = c 2 ∑ k = 1 K a k 2 , H 1 ( J β K ∩ Q ) = c J ∑ k = 1 K r k , ∫ J β K ∣ β K + − β K − ∣ d H 1 = c D ∑ k = 1 K a k r k . \|\nabla\beta_{K}\|_{L^{2}(Q)}^{2}=c_{2}\sum_{k=1}^{K}a_{k}^{2},\qquad\mathcal{H}^{1}(J_{\beta_{K}}\cap Q)=c_{J}\sum_{k=1}^{K}r_{k},\qquad\int_{J_{\beta_{K}}}|\beta_{K}^{+}-\beta_{K}^{-}|\,\mathrm{d}\mathcal{H}^{1}=c_{D}\sum_{k=1}^{K}a_{k}r_{k}. ∥∇ β K ∥ L 2 ( Q ) 2 = c 2 k = 1 ∑ K a k 2 , H 1 ( J β K ∩ Q ) = c J k = 1 ∑ K r k , ∫ J β K ∣ β K + − β K − ∣ d H 1 = c D k = 1 ∑ K a k r k .
Obviously r k → 0 r_k \to0 r k → 0 ; we select a k = 2 − k a_k=2^{-k} a k = 2 − k . Then the sequence ( β K ) K (\beta_K)_K ( β K ) K converges uniformly and strongly in B V BV B V to some β ∞ \beta_\infty β ∞ . In each ball B k B_k B k , we have β ∞ = β h \beta_\infty=\beta_h β ∞ = β h for any h ≥ k h\geq k h ≥ k ; outside the union of the balls β ∞ = 0 \beta_\infty=0 β ∞ = 0 . One checks that u : = [ β ∞ ] ∈ H 1 ( Q ; M d × d / G ) u:=[\beta_\infty]\in H^1(Q;\mathbb{M}^{d\times d}/G) u := [ β ∞ ] ∈ H 1 ( Q ; M d × d / G ) .
It remains to show that there is no β ∈ S B V G ( Q ; M d × d ) \beta\in SBV_G(Q;\mathbb{M}^{d\times d}) β ∈ S B V G ( Q ; M d × d ) with [ β ] = u [\beta]=u [ β ] = u .
Assume there was one, pick any h ∈ N h\in\mathbb{N} h ∈ N , and consider the ball B h B_h B h . We claim that necessarily
H 1 ( J β ∩ B h ) ≥ 1 2 r h . (36) \mathcal{H}^1(J_\beta\cap B_h)\geq\frac{1}{2}r_h. \tag*{(36)} H 1 ( J β ∩ B h ) ≥ 2 1 r h . (36)
If this holds, then H 1 ( J β ∩ Q ) ≥ 1 2 ∑ h r h = ∞ \mathcal{H}^1(J_\beta\cap Q)\geq\frac{1}{2}\sum_h r_h=\infty H 1 ( J β ∩ Q ) ≥ 2 1 ∑ h r h = ∞ , so that β ∉ S B V 2 ( Q ; M d × d ) \beta\notin SBV^2(Q;\mathbb{M}^{d\times d}) β ∈ / S B V 2 ( Q ; M d × d ) and the proof is concluded.
It remains to prove (2.15). Assume this was not the case. Then by slicing there would be a set E ⊂ ( 0 , 1 ) E\subset(0,1) E ⊂ ( 0 , 1 ) with L 1 ( E ) > 1 / 2 \mathcal{L}^1(E)>1/2 L 1 ( E ) > 1/2 such that for any t ∈ E t\in E t ∈ E the 2 π 2\pi 2 π -periodic function
v ( θ ) : = β ( x h + r h t e i θ ) v(\theta):=\beta(x_h+r_h t e^{i\theta}) v ( θ ) := β ( x h + r h t e i θ )
obeys v ∈ W l o c 1 , 1 ( R ; M d × d ) v\in W_{\mathrm{loc}}^{1,1}(\mathbb{R};\mathbb{M}^{d\times d}) v ∈ W loc 1 , 1 ( R ; M d × d ) , we identify it with its continuous representative. By our definitions, [ v ( θ ) ] = a h φ ( t ) [ γ ( θ ) ] [v(\theta)]=a_h\varphi(t)[\gamma(\theta)] [ v ( θ )] = a h φ ( t ) [ γ ( θ )] . We pick t ∈ E t\in E t ∈ E with φ ( t ) > 0 \varphi(t)>0 φ ( t ) > 0 and seek a contradiction. We define w : [ 0 , 2 π ] → M d × d w:[0,2\pi]\to\mathbb{M}^{d\times d} w : [ 0 , 2 π ] → M d × d by
w ( θ ) : = v ( θ ) ( γ ( θ ) ) − 1 . w(\theta):=v(\theta)(\gamma(\theta))^{-1}. w ( θ ) := v ( θ ) ( γ ( θ ) ) − 1 .
This function takes values in a h φ ( t ) G a_h\varphi(t)G a h φ ( t ) G . We recall that a h φ ( t ) ≠ 0 a_h\varphi(t)\neq0 a h φ ( t ) = 0 , so that this set consists of invertible matrices. Using periodicity of v v v and the boundary conditions on γ \gamma γ we obtain
w ( 0 ) = v ( 0 ) ( γ ( 0 ) ) − 1 = v ( 0 ) , w ( 2 π ) = v ( 2 π ) ( γ ( 2 π ) ) − 1 = v ( 0 ) G − 1 . w(0)=v(0)(\gamma(0))^{-1}=v(0),\qquad w(2\pi)=v(2\pi)(\gamma(2\pi))^{-1}=v(0)G^{-1}. w ( 0 ) = v ( 0 ) ( γ ( 0 ) ) − 1 = v ( 0 ) , w ( 2 π ) = v ( 2 π ) ( γ ( 2 π ) ) − 1 = v ( 0 ) G − 1 .
However, as w w w is continuous on [ 0 , 2 π ] [0,2\pi] [ 0 , 2 π ] , it is constant. Since v ( 0 ) v(0) v ( 0 ) is an invertible matrix, this gives G = Id G=\operatorname{Id} G = Id , a contradiction. □
Functional setting II: Metric-space valued BV functions
We recall from [1 ] the definition of functions of Bounded Variation into a general separable locally compact metric space ( X , d X ) (X,d_X) ( X , d X ) .
Definition 2.7. Let u ∈ L 1 ( Ω ; X ) u\in L^1(\Omega;X) u ∈ L 1 ( Ω ; X ) . We say that u u u belongs to B V ( Ω ; X ) BV(\Omega;X) B V ( Ω ; X ) if for every φ ∈ Lip ( X ; R m ) \varphi\in\operatorname{Lip}(X;\mathbb{R}^m) φ ∈ Lip ( X ; R m ) one has φ ∘ u ∈ B V ( Ω ; R m ) \varphi\circ u\in BV(\Omega;\mathbb{R}^m) φ ∘ u ∈ B V ( Ω ; R m ) and there exists a finite Borel measure σ \sigma σ (which may depend on m m m , but not on φ \varphi φ ) such that
∣ D ( φ ∘ u ) ∣ ≤ Lip ( φ ) σ . (37) |D(\varphi\circ u)|\leq\operatorname{Lip}(\varphi)\,\sigma. \tag*{(37)} ∣ D ( φ ∘ u ) ∣ ≤ Lip ( φ ) σ . (37)
In this case, for m ∈ N > 0 m\in\mathbb{N}_{>0} m ∈ N > 0 the total variation measure of u u u is defined as
∣ D u ∣ ( m ) : = sup { ∣ D ( φ ∘ u ) ∣ : φ ∈ Lip 1 ( X ; R m ) } , (38) |Du|_{(m)}:=\sup\{|D(\varphi\circ u)|:\varphi\in\operatorname{Lip}_1(X;\mathbb{R}^m)\}, \tag*{(38)} ∣ D u ∣ ( m ) := sup { ∣ D ( φ ∘ u ) ∣ : φ ∈ Lip 1 ( X ; R m )} , (38)
where the supremum of measures is taken as in Definition 1.68 in [3 ] .
As for the case of Sobolev functions, the value of ∣ D u ∣ ( m ) |Du|_{(m)} ∣ D u ∣ ( m ) depends on m m m (see Remark 2.3). The usual decomposition of ∣ D u ∣ |Du| ∣ D u ∣ holds true in B V ( Ω ; X ) BV(\Omega;X) B V ( Ω ; X ) (see Theorems 2.2 and 2.3, and Remark 2.3 in [1 ] ):
|Du|_{(m)}=|\nabla u|_{(m)}\mathcal{L}^N+d_X(u^+,u^-)\mathcal{H}^{N-1}\mathbin{\vrule height 1.4ex depth -0.4ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.7ex}J_u+|Du|^c_{(m)}, \tag*{(39)}
where ∣ ∇ u ∣ ( m ) ( x ) = sup { ∣ ∇ ( φ ∘ u ) ∣ ( x ) : φ ∈ Lip 1 ( X ; R m ) } ∈ L 1 ( Ω ) |\nabla u|_{(m)}(x)=\sup\{|\nabla(\varphi\circ u)|(x):\varphi\in\operatorname{Lip}_1(X;\mathbb{R}^m)\}\in L^1(\Omega) ∣∇ u ∣ ( m ) ( x ) = sup { ∣∇ ( φ ∘ u ) ∣ ( x ) : φ ∈ Lip 1 ( X ; R m )} ∈ L 1 ( Ω ) , J u J_u J u is a countably H N − 1 \mathcal{H}^{N-1} H N − 1 -rectifiable subset of R N \mathbb{R}^N R N with unit normal ν ( x ) \nu(x) ν ( x ) defined for H N − 1 \mathcal{H}^{N-1} H N − 1 -a.e. x ∈ J u x\in J_u x ∈ J u , and the traces u ± ( x ) = u ± ( x , ν ( x ) ) u^\pm(x)=u^\pm(x,\nu(x)) u ± ( x ) = u ± ( x , ν ( x )) are characterized by
lim ρ → 0 + L N ( B ρ ( x ) ∩ { y ∈ Ω : ± ( y − x ) ⋅ ν ( x ) > 0 , d X ( u ( y ) , u ± ( x ) ) > ε } ) ρ d = 0 ∀ ε > 0. (40) \lim_{\rho\to0^+}\frac{\mathcal{L}^N\left(B_\rho(x)\cap\{y\in\Omega:\ \pm(y-x)\cdot\nu(x)>0,\ d_X(u(y),u^\pm(x))>\varepsilon\}\right)}{\rho^d}=0\qquad\forall\ \varepsilon>0. \tag*{(40)} ρ → 0 + lim ρ d L N ( B ρ ( x ) ∩ { y ∈ Ω : ± ( y − x ) ⋅ ν ( x ) > 0 , d X ( u ( y ) , u ± ( x )) > ε } ) = 0 ∀ ε > 0. (40)
Remark 2.8. Following [1 ] , u ∈ S B V ( Ω ; X ) u \in\mathrm{SBV}(\Omega;X) u ∈ SBV ( Ω ; X ) if u ∈ B V ( Ω ; X ) u \in\mathrm{BV}(\Omega;X) u ∈ BV ( Ω ; X ) and ∣ D u ∣ ( m ) c = 0 |Du|_{(m)}^{c}=0 ∣ D u ∣ ( m ) c = 0 . This is equivalent to require that u ∈ B V ( Ω ; X ) u \in\mathrm{BV}(\Omega;X) u ∈ BV ( Ω ; X ) and φ ∘ u ∈ S B V ( Ω ; R m ) \varphi\circ u \in\mathrm{SBV}(\Omega;\mathbb{R}^{m}) φ ∘ u ∈ SBV ( Ω ; R m ) for every φ ∈ Lip ( X ; R m ) \varphi\in\operatorname{Lip}(X;\mathbb{R}^{m}) φ ∈ Lip ( X ; R m ) (see [1 Remark 2.4] ).
As in the case of X X X -valued Sobolev functions, the norm depends on m m m , but the space does not. In working with M d × d / G \mathbb{M}^{d \times d}/\mathcal{G} M d × d / G we shall mostly take m = d 2 m=d^{2} m = d 2 and, when there is no risk of confusion, simply write ∣ D u ∣ |Du| ∣ D u ∣ for ∣ D u ∣ ( d 2 ) |Du|_{(d^{2})} ∣ D u ∣ ( d 2 ) .
Remark 2.9. The decomposition (39) and definition (38) also allow to characterize subspaces of B V ( Ω ; X ) \mathrm{BV}(\Omega;X) BV ( Ω ; X ) .
To this aim we recall the following basic property of the derivation of measures, also used in [1 ] : if σ \sigma σ is a positive measure, and μ \mu μ , with μ ( Ω ) < + ∞ \mu(\Omega)<+\infty μ ( Ω ) < + ∞ , is the supremum of a countable family of positive measures μ h \mu_h μ h , and μ h = f h σ + μ h s \mu_h=f_h\sigma+\mu_h^{s} μ h = f h σ + μ h s , where μ h s \mu_h^{s} μ h s denotes the singular part of μ h \mu_h μ h with respect to σ \sigma σ and f h ∈ L 1 ( Ω ; σ ) f_h \in L^{1}(\Omega;\sigma) f h ∈ L 1 ( Ω ; σ ) (i.e. f h f_h f h is in L 1 L^{1} L 1 with respect to the measure σ \sigma σ ), then
μ = f σ + μ s , with f = sup f h σ -a.e. in Ω . (41) \mu=f\sigma+\mu^{s}, \qquad\text{with } f=\sup f_h \quad\sigma\text{-a.e. in }\Omega. \tag*{(41)} μ = f σ + μ s , with f = sup f h σ -a.e. in Ω. (41)
Therefore from Remark 2.8 we have
u ∈ S B V ( Ω ; X ) ⟺ φ ∘ u ∈ S B V ( Ω ; R m ) for all φ ∈ Lip ( X ; R m ) . (42) u \in\mathrm{SBV}(\Omega;X) \Longleftrightarrow\varphi\circ u \in\mathrm{SBV}(\Omega;\mathbb{R}^{m}) \text{ for all } \varphi\in\operatorname{Lip}(X;\mathbb{R}^{m}). \tag*{(42)} u ∈ SBV ( Ω ; X ) ⟺ φ ∘ u ∈ SBV ( Ω ; R m ) for all φ ∈ Lip ( X ; R m ) . (42)
Moreover we also define the space S B V 0 ( Ω ; X ) \mathrm{SBV}_{0}(\Omega;X) SBV 0 ( Ω ; X ) as the set of functions u ∈ S B V ( Ω ; X ) u \in\mathrm{SBV}(\Omega;X) u ∈ SBV ( Ω ; X ) such that ∣ ∇ u ∣ = 0 |\nabla u|=0 ∣∇ u ∣ = 0 , which is also equivalent to
φ ∘ u ∈ S B V 0 ( Ω ; R m ) for all φ ∈ Lip ( X ; R m ) . (43) \varphi\circ u \in\mathrm{SBV}_{0}(\Omega;\mathbb{R}^{m}) \text{ for all } \varphi\in\operatorname{Lip}(X;\mathbb{R}^{m}). \tag*{(43)} φ ∘ u ∈ SBV 0 ( Ω ; R m ) for all φ ∈ Lip ( X ; R m ) . (43)
Finally, when u ∈ S B V ( Ω ; X ) u \in\mathrm{SBV}(\Omega;X) u ∈ SBV ( Ω ; X ) has no jump, and ∣ ∇ u ∣ ( m ) ∈ L p ( Ω ) |\nabla u|_{(m)} \in L^{p}(\Omega) ∣∇ u ∣ ( m ) ∈ L p ( Ω ) , the notion above complies with the definition of W 1 , p ( Ω ; X ) W^{1,p}(\Omega;X) W 1 , p ( Ω ; X ) in [36 ] recalled in Definition 2.2.
We recall here the compactness results in BV proven in [1 Theorem 2.4] .
Theorem 2.10. Let ( u n ) n ∈ N (u_n)_{n\in\mathbb{N}} ( u n ) n ∈ N be a sequence of functions in B V ( Ω ; X ) \mathrm{BV}(\Omega;X) BV ( Ω ; X ) , and m ∈ N > 0 m\in\mathbb{N}_{>0} m ∈ N > 0 . If there exists z 0 ∈ X z_0\in X z 0 ∈ X such that
sup n ∈ N ( ∣ D u n ∣ ( m ) ( Ω ) + ∫ Ω d X ( u n ( x ) , z 0 ) d x ) < + ∞ , \sup_{n\in\mathbb{N}}\left(|Du_n|_{(m)}(\Omega)+\int_{\Omega}d_X(u_n(x),z_0)\,\mathrm{d}x\right)<+\infty, n ∈ N sup ( ∣ D u n ∣ ( m ) ( Ω ) + ∫ Ω d X ( u n ( x ) , z 0 ) d x ) < + ∞ , then there exist a subsequence ( u n k ) k (u_{n_k})_k ( u n k ) k and a function u ∈ B V ( Ω ; X ) u\in\mathrm{BV}(\Omega;X) u ∈ BV ( Ω ; X ) such that u n k u_{n_k} u n k converges to u u u a.e. in Ω \Omega Ω , and
∣ D u ∣ ( m ) ( Ω ) ≤ lim inf k → + ∞ ∣ D u n k ∣ ( m ) ( Ω ) . |Du|_{(m)}(\Omega)\leq\liminf_{k\to+\infty}|Du_{n_k}|_{(m)}(\Omega). ∣ D u ∣ ( m ) ( Ω ) ≤ k → + ∞ lim inf ∣ D u n k ∣ ( m ) ( Ω ) .
We also recall that one can characterize functions in B V ( Ω ; X ) \mathrm{BV}(\Omega;X) BV ( Ω ; X ) via slicing as follows. We fix ξ ∈ S N − 1 : = { ξ ∈ R N : ∣ ξ ∣ = 1 } \xi\in\mathbb{S}^{N-1}:=\{\xi\in\mathbb{R}^{N}:|\xi|=1\} ξ ∈ S N − 1 := { ξ ∈ R N : ∣ ξ ∣ = 1 } and we set
Π ξ : = { y ∈ R N : y ⋅ ξ = 0 } , Ω y ξ : = { t ∈ R : y + t ξ ∈ Ω } , Ω ξ : = { y ∈ Π ξ : Ω y ξ ≠ ∅ } . \Pi^{\xi}:=\{y\in\mathbb{R}^{N}:y\cdot\xi=0\},\qquad\Omega_{y}^{\xi}:=\{t\in\mathbb{R}:y+t\xi\in\Omega\},\qquad\Omega^{\xi}:=\{y\in\Pi^{\xi}:\Omega_{y}^{\xi}\ne\varnothing\}. Π ξ := { y ∈ R N : y ⋅ ξ = 0 } , Ω y ξ := { t ∈ R : y + t ξ ∈ Ω } , Ω ξ := { y ∈ Π ξ : Ω y ξ = ∅ } .
For any function u : Ω → X u:\Omega\to X u : Ω → X we define the slices u y ξ : Ω y ξ → X u_{y}^{\xi}:\Omega_{y}^{\xi}\to X u y ξ : Ω y ξ → X by
u y ξ ( t ) : = u ( y + t ξ ) . u_{y}^{\xi}(t):=u(y+t\xi). u y ξ ( t ) := u ( y + t ξ ) .
The following result is an immediate consequence of [1 Proposition 2.1] , stated with our notation and also formulated for Sobolev functions.
Proposition 2.11. The following hold:
if u ∈ B V ( Ω ; X ) u\in\mathrm{BV}(\Omega;X) u ∈ BV ( Ω ; X ) , then for every ξ ∈ S N − 1 \xi\in\mathbb{S}^{N-1} ξ ∈ S N − 1 we have u y ξ ∈ B V ( Ω y ξ ; X ) u_{y}^{\xi}\in\mathrm{BV}(\Omega_{y}^{\xi};X) u y ξ ∈ BV ( Ω y ξ ; X ) and ∣ D u ∣ ( 1 ) ≥ ∣ D ( u y ξ ) ∣ |Du|_{(1)}\geq|D(u_{y}^{\xi})| ∣ D u ∣ ( 1 ) ≥ ∣ D ( u y ξ ) ∣ as measures in Ω y ξ \Omega_{y}^{\xi} Ω y ξ , for H N − 1 \mathcal{H}^{N-1} H N − 1 -a.e. y ∈ Ω ξ y\in\Omega^{\xi} y ∈ Ω ξ ; (2.23)
ii) If u ∈ W 1 , p ( Ω ; X ) u \in W^{1,p}(\Omega; X) u ∈ W 1 , p ( Ω ; X ) , then for every ξ ∈ S N − 1 \xi\in S^{N-1} ξ ∈ S N − 1 we have u y ξ ∈ W 1 , p ( Ω y ξ ; X ) u_y^\xi\in W^{1,p}(\Omega_y^\xi; X) u y ξ ∈ W 1 , p ( Ω y ξ ; X ) and
∣ ∇ u ( x ) ∣ ≥ ∣ ( u y ξ ) ′ ( t x ) ∣ for H N − 1 -a.e. y ∈ Ω ξ and a.e. x ∈ Ω , |\nabla u(x)| \ge|(u_y^\xi)'(t_x)| \quad\text{for } \mathcal{H}^{N-1}\text{-a.e. } y \in\Omega^\xi\text{ and a.e. } x \in\Omega, ∣∇ u ( x ) ∣ ≥ ∣ ( u y ξ ) ′ ( t x ) ∣ for H N − 1 -a.e. y ∈ Ω ξ and a.e. x ∈ Ω ,
where t x ∈ Ω y ξ t_x \in\Omega_y^\xi t x ∈ Ω y ξ is such that x = y + t x ξ x = y + t_x \xi x = y + t x ξ .
Proof. The proof of i) follows from [1 Proposition 2.1] and a localization argument. Property ii) follows from i) and the characterization of ∣ D u ∣ |Du| ∣ D u ∣ given in (2.18). □
We prove that maps in SBV ( Ω ; O ( d ) / G ) \operatorname{SBV}(\Omega; \mathrm{O}(d)/G) SBV ( Ω ; O ( d ) / G ) have a lifting in SBV ( Ω ; O ( d ) ) \operatorname{SBV}(\Omega; \mathrm{O}(d)) SBV ( Ω ; O ( d )) , with a bound on the norm. This could be obtained easily as a consequence of a general result by Canevari and Orlandi [14 ] on maps valued in compact manifolds with a universal covering. In our case, the manifold is O ( d ) / G \mathrm{O}(d)/G O ( d ) / G , and permits a simpler argument. We remark that in the case d = 2 d=2 d = 2 , this also follows from the simple proof in [23 ] for maps valued in S 1 S^1 S 1 . In higher dimension, however, a localization is needed, as in [14 ] . Our proof applies without significant changes both for S O ( d ) \mathrm{SO}(d) SO ( d ) and for O ( d ) \mathrm{O}(d) O ( d ) ; we present it only in the case of O ( d ) \mathrm{O}(d) O ( d ) , which is the one used below.
Theorem 2.12. Let u ∈ SBV ( Ω ; O ( d ) / G ) u \in\operatorname{SBV}(\Omega; \mathrm{O}(d)/G) u ∈ SBV ( Ω ; O ( d ) / G ) , Ω ⊂ R N \Omega\subset\mathbb{R}^{N} Ω ⊂ R N open and bounded. Then there is β ∈ SBV ( Ω ; O ( d ) ) \beta\in\operatorname{SBV}(\Omega; \mathrm{O}(d)) β ∈ SBV ( Ω ; O ( d )) such that
u = [ β ] pointwise a.e. in Ω , and ∣ D β ∣ ( Ω ) ≤ C ∣ D u ∣ ( d 2 ) ( Ω ) . (44) u = [\beta] \text{ pointwise a.e. in } \Omega, \text{ and } |D\beta|(\Omega) \le C|Du|^{(d^2)}(\Omega). \tag*{(44)} u = [ β ] pointwise a.e. in Ω , and ∣ D β ∣ ( Ω ) ≤ C ∣ D u ∣ ( d 2 ) ( Ω ) . (44) The constant depends only on d d d and G G G .
Proof. Let r 0 : = 1 4 min { ∣ G − G ′ ∣ : G , G ′ ∈ G , G ≠ G ′ } r_0 := \frac{1}{4}\min\{|G-G'|: G,G' \in G, G \ne G'\} r 0 := 4 1 min { ∣ G − G ′ ∣ : G , G ′ ∈ G , G = G ′ } .
Step 1. We can identify B r 0 ( [ Z ] ) B_{r_0}([Z]) B r 0 ([ Z ]) with B r 0 ( Z ) B_{r_0}(Z) B r 0 ( Z ) . Fix Z ∈ O ( d ) / G Z \in\mathrm{O}(d)/G Z ∈ O ( d ) / G . We check that Y ↦ [ Y ] Y \mapsto[Y] Y ↦ [ Y ] is an isometric bijection of B r 0 ( Z ) ⊂ O ( d ) B_{r_0}(Z) \subset\mathrm{O}(d) B r 0 ( Z ) ⊂ O ( d ) onto B r 0 ( [ Z ] ) ⊂ O ( d ) / G B_{r_0}([Z]) \subset\mathrm{O}(d)/G B r 0 ([ Z ]) ⊂ O ( d ) / G . To see this, the key fact is that if Y , Y ′ ∈ B r 0 ( Z ) Y,Y' \in B_{r_0}(Z) Y , Y ′ ∈ B r 0 ( Z ) and G ∈ G ∖ { Id } G \in\mathcal{G} \setminus\{\operatorname{Id}\} G ∈ G ∖ { Id } then
∣ G Y − Y ′ ∣ > ∣ Y − Y ′ ∣ , (45) |GY-Y'|>|Y-Y'|, \tag*{(45)} ∣ G Y − Y ′ ∣ > ∣ Y − Y ′ ∣ , (45)
which implies that d G ( [ Y ] , [ Y ′ ] ) = ∣ Y − Y ′ ∣ d_G([Y],[Y'])=|Y-Y'| d G ([ Y ] , [ Y ′ ]) = ∣ Y − Y ′ ∣ . In turn, to show (2.25) it suffices to observe that, since Y ∈ O ( d ) Y \in\mathrm{O}(d) Y ∈ O ( d ) and ∣ Y − Y ′ ∣ ≤ ∣ Y − Z ∣ + ∣ Y ′ − Z ∣ < 2 r 0 |Y-Y'| \le|Y-Z|+|Y'-Z|<2r_0 ∣ Y − Y ′ ∣ ≤ ∣ Y − Z ∣ + ∣ Y ′ − Z ∣ < 2 r 0 ,
∣ G Y − Y ′ ∣ = ∣ ( G − Id ) Y + ( Y − Y ′ ) ∣ ≥ ∣ G − Id ∣ − ∣ Y − Y ′ ∣ ≥ 4 r 0 − 2 r 0 > ∣ Y − Y ′ ∣ . (46) |GY-Y'|=|(G-\operatorname{Id})Y+(Y-Y')| \ge|G-\operatorname{Id}|-|Y-Y'| \ge4r_0-2r_0>|Y-Y'|. \tag*{(46)} ∣ G Y − Y ′ ∣ = ∣ ( G − Id ) Y + ( Y − Y ′ ) ∣ ≥ ∣ G − Id ∣ − ∣ Y − Y ′ ∣ ≥ 4 r 0 − 2 r 0 > ∣ Y − Y ′ ∣. (46)
We denote by ψ Z : B r 0 ( [ Z ] ) → B r 0 ( Z ) ⊂ O ( d ) \psi_Z:B_{r_0}([Z])\to B_{r_0}(Z)\subset\mathrm{O}(d) ψ Z : B r 0 ([ Z ]) → B r 0 ( Z ) ⊂ O ( d ) the inverse of Y ↦ [ Y ] Y\mapsto[Y] Y ↦ [ Y ] .
Step 2. Localization via the coarea formula. We claim that for any Z ∈ O ( d ) Z \in\mathrm{O}(d) Z ∈ O ( d ) there is r ∈ ( r 0 / 2 , r 0 ) r \in(r_0/2,r_0) r ∈ ( r 0 /2 , r 0 ) such that
H N − 1 ( ∂ ∗ ( u − 1 ( B r ( [ Z ] ) ) ) ) < c ∣ D u ∣ ( Ω ) . (47) \mathcal{H}^{N-1}(\partial^*(u^{-1}(B_r([Z]))))<c|Du|(\Omega). \tag*{(47)} H N − 1 ( ∂ ∗ ( u − 1 ( B r ([ Z ])))) < c ∣ D u ∣ ( Ω ) . (47)
To see this, consider φ : O ( d ) / G → R \varphi:\mathrm{O}(d)/G\to\mathbb{R} φ : O ( d ) / G → R defined by φ ( Y ) : = d G ( Y , [ Z ] ) \varphi(Y):=d_G(Y,[Z]) φ ( Y ) := d G ( Y , [ Z ]) , and notice that u − 1 ( B t ( [ Z ] ) ) = { x : φ ( u ( x ) ) < t } u^{-1}(B_t([Z]))=\{x:\varphi(u(x))<t\} u − 1 ( B t ([ Z ])) = { x : φ ( u ( x )) < t } . The function φ \varphi φ is obviously 1 1 1 -Lipschitz, therefore φ ∘ u ∈ SBV ( Ω ) \varphi\circ u\in\operatorname{SBV}(\Omega) φ ∘ u ∈ SBV ( Ω ) , and ∣ D ( φ ∘ u ) ∣ ( Ω ) ≤ ∣ D u ∣ ( Ω ) |D(\varphi\circ u)|(\Omega)\le|Du|(\Omega) ∣ D ( φ ∘ u ) ∣ ( Ω ) ≤ ∣ D u ∣ ( Ω ) . By the coarea formula applied to the SBV function φ ∘ u : Ω → R \varphi\circ u:\Omega\to\mathbb{R} φ ∘ u : Ω → R we obtain
∫ R H N − 1 ( ∂ ∗ { φ ∘ u < t } ) d t = ∣ D ( φ ∘ u ) ∣ ( Ω ) , (48) \int_{\mathbb{R}}\mathcal{H}^{N-1}(\partial^*\{\varphi\circ u<t\})\,dt=|D(\varphi\circ u)|(\Omega), \tag*{(48)} ∫ R H N − 1 ( ∂ ∗ { φ ∘ u < t }) d t = ∣ D ( φ ∘ u ) ∣ ( Ω ) , (48)
therefore we can choose r r r as in (2.27), with c = 2 / r 0 c=2/r_0 c = 2/ r 0 depending only on G G G .
Step 3. Decomposition of the domain. Let Z 1 , … , Z M ∈ O ( d ) Z_1,\ldots,Z_M\in\mathrm{O}(d) Z 1 , … , Z M ∈ O ( d ) be such that O ( d ) / G ⊂ ⋃ i B r 0 / 2 ( [ Z i ] ) \mathrm{O}(d)/G\subset\bigcup_i B_{r_0/2}([Z_i]) O ( d ) / G ⊂ ⋃ i B r 0 /2 ([ Z i ]) (this choice depends only on G G G ), and let r i r_i r i be the corresponding values with the property (2.27) (which depend also on u u u ). We define iteratively
ω 1 : = u − 1 ( B r 1 ( [ Z 1 ] ) ) , ω i : = u − 1 ( B r i ( [ Z i ] ) ) ∖ ⋃ j < i ω j . \omega_1:=u^{-1}(B_{r_1}([Z_1])), \qquad\omega_i:=u^{-1}(B_{r_i}([Z_i]))\setminus\bigcup_{j<i}\omega_j. ω 1 := u − 1 ( B r 1 ([ Z 1 ])) , ω i := u − 1 ( B r i ([ Z i ])) ∖ j < i ⋃ ω j .
The sets are a partition of Ω \Omega Ω , and H N − 1 ( ∂ ∗ ω i ) ≤ c ∣ D u ∣ ( Ω ) \mathcal{H}^{N-1}(\partial^*\omega_i)\le c|Du|(\Omega) H N − 1 ( ∂ ∗ ω i ) ≤ c ∣ D u ∣ ( Ω ) for each i i i , with c c c depending only on the group G G G . Step 4. Conclusion. We set β : = ψ Z i ∘ u \beta:= \psi_{Z_i} \circ u β := ψ Z i ∘ u in each ω i \omega_i ω i . Obviously [ β ] = u [\beta] = u [ β ] = u , and inside each ω i \omega_i ω i we have ∣ D β ∣ ≤ C ∣ D u ∣ ( d 2 ) |D\beta| \leq C|Du|_{(d^2)} ∣ D β ∣ ≤ C ∣ D u ∣ ( d 2 ) as measures. We compute
∣ D β ∣ ( Ω ) ≤ ∑ i = 1 M ∣ D β ∣ ( ω i ) + ∑ i = 1 M ∥ β ∥ L ∞ H N − 1 ( ∂ ∗ ω i ) ≤ C ∣ D u ∣ ( d 2 ) ( Ω ) , (49) |D\beta|(\Omega) \leq\sum_{i=1}^{M} |D\beta|(\omega_i) + \sum_{i=1}^{M} \|\beta\|_{L^\infty} \mathcal{H}^{N-1}(\partial^*\omega_i) \leq C|Du|_{(d^2)}(\Omega), \tag*{(49)} ∣ D β ∣ ( Ω ) ≤ i = 1 ∑ M ∣ D β ∣ ( ω i ) + i = 1 ∑ M ∥ β ∥ L ∞ H N − 1 ( ∂ ∗ ω i ) ≤ C ∣ D u ∣ ( d 2 ) ( Ω ) , (49)
again with C C C depending only on G \mathcal{G} G and d d d .
Remark 2.13. By a localization argument, it is simple to see that if u ∈ SBV 0 ( Ω ; O ( d ) / G ) u \in\operatorname{SBV}_0(\Omega; \mathrm{O}(d)/\mathcal{G}) u ∈ SBV 0 ( Ω ; O ( d ) / G ) , then the function β \beta β whose existence is guaranteed by Theorem 2.12 satisfies β ∈ SBV 0 ( Ω ; O ( d ) ) \beta\in\operatorname{SBV}_0(\Omega; \mathrm{O}(d)) β ∈ SBV 0 ( Ω ; O ( d )) . Moreover, if any other β ∗ \beta_* β ∗ with this property exists, then β ∗ = G β \beta_* = G\beta β ∗ = Gβ for some G ∈ BV ( Ω ; G ) G \in\operatorname{BV}(\Omega; \mathcal{G}) G ∈ BV ( Ω ; G ) .
Remark 2.14. An extension of the argument used in Theorem 2.12 to construct a lifting of a map in SBV ( Ω ; O ( d ) / G ) \operatorname{SBV}(\Omega; \mathrm{O}(d)/\mathcal{G}) SBV ( Ω ; O ( d ) / G ) to a map in SBV G ( Ω ; O ( d ) ) \operatorname{SBV}_{\mathcal{G}}(\Omega; \mathrm{O}(d)) SBV G ( Ω ; O ( d )) permits to show that if G ⊂ SO ( 2 ) \mathcal{G} \subset\operatorname{SO}(2) G ⊂ SO ( 2 ) then any map in SBV ( Ω ; M 2 × 2 / G ) \operatorname{SBV}(\Omega; \mathbb{M}^{2\times2}/\mathcal{G}) SBV ( Ω ; M 2 × 2 / G ) admits a representative in SBV ( Ω ; M 2 × 2 ) \operatorname{SBV}(\Omega; \mathbb{M}^{2\times2}) SBV ( Ω ; M 2 × 2 ) . The treatment of the higher-dimensional case is more complex and outside of the scope of this paper. Therefore we postpone a more detailed analysis of the relation between the spaces SBV ( Ω ; M d × d / G ) \operatorname{SBV}(\Omega; \mathbb{M}^{d\times d}/\mathcal{G}) SBV ( Ω ; M d × d / G ) and SBV ( Ω ; M d × d ) \operatorname{SBV}(\Omega; \mathbb{M}^{d\times d}) SBV ( Ω ; M d × d ) to future work.
Metric-space valued phase field models
With the above functional framework at hand, we can formulate the approximating phase field model in terms of order parameters valued into equivalence classes. The natural domain will be given by Sobolev functions with values in M d × d / G \mathbb{M}^{d\times d}/\mathcal{G} M d × d / G . We define
F ~ ε ( u , v ; A ) : = ∫ A f ε 2 ( v ) ∣ ∇ u ∣ ( d 2 ) 2 d x + P F ε ( v ; A ) + 1 δ ε ∫ A d G 2 ( u , O ( d ) / G ) d x (50) \widetilde{\mathcal{F}}_\varepsilon(u,v;A) := \int_A f_\varepsilon^2(v)|\nabla u|_{(d^2)}^2\,dx + \operatorname{P}\mathcal{F}_\varepsilon(v;A) + \frac{1}{\delta_\varepsilon}\int_A d_{\mathcal{G}}^2(u,\mathrm{O}(d)/\mathcal{G})\,dx \tag*{(50)} F ε ( u , v ; A ) := ∫ A f ε 2 ( v ) ∣∇ u ∣ ( d 2 ) 2 d x + P F ε ( v ; A ) + δ ε 1 ∫ A d G 2 ( u , O ( d ) / G ) d x (50)
for A ⊂ Ω A \subset\Omega A ⊂ Ω open and
( u , v ) ∈ H 1 ( Ω ; M d × d / G ) × H 1 ( Ω ; [ 0 , 1 ] ) , (u,v) \in H^1(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}) \times H^1(\Omega;[0,1]), ( u , v ) ∈ H 1 ( Ω ; M d × d / G ) × H 1 ( Ω ; [ 0 , 1 ]) ,
and equal to + ∞ +\infty + ∞ otherwise in L 1 ( Ω ; M d × d / G ) × L 1 ( Ω ) L^1(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}) \times L^1(\Omega) L 1 ( Ω ; M d × d / G ) × L 1 ( Ω ) .
By Remark 2.4, any β ∈ SBV G ( Ω ; M d × d ) \beta\in\operatorname{SBV}_{\mathcal{G}}(\Omega;\mathbb{M}^{d\times d}) β ∈ SBV G ( Ω ; M d × d ) is such that [ β ] ∈ H 1 ( Ω ; M d × d / G ) [\beta] \in H^1(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}) [ β ] ∈ H 1 ( Ω ; M d × d / G ) with ∣ ∇ [ β ] ∣ ( d 2 ) ≤ ∣ ∇ β ∣ |\nabla[\beta]|_{(d^2)} \leq|\nabla\beta| ∣∇ [ β ] ∣ ( d 2 ) ≤ ∣∇ β ∣ , and obviously d G ( [ β ] , O ( d ) / G ) = dist ( β , O ( d ) ) d_{\mathcal{G}}([\beta],\mathrm{O}(d)/\mathcal{G}) = \operatorname{dist}(\beta,\mathrm{O}(d)) d G ([ β ] , O ( d ) / G ) = dist ( β , O ( d )) . It follows that
F ~ ε ( [ β ] , v ; A ) ≤ F ε ( β , v ; A ) . (51) \widetilde{\mathcal{F}}_\varepsilon([\beta],v;A) \leq\mathcal{F}_\varepsilon(\beta,v;A). \tag*{(51)} F ε ([ β ] , v ; A ) ≤ F ε ( β , v ; A ) . (51)
Finally in view of the invariance (1.14) and Remark 2.13 we can define
F ~ λ ( u , 1 ) : = F λ ( β , 1 ) for every u ∈ SBV 0 ( Ω ; O ( d ) / G ) (52) \widetilde{\mathcal{F}}^\lambda(u,1) := \mathcal{F}^\lambda(\beta,1) \quad\text{for every } u \in\operatorname{SBV}_0(\Omega;\mathrm{O}(d)/\mathcal{G}) \tag*{(52)} F λ ( u , 1 ) := F λ ( β , 1 ) for every u ∈ SBV 0 ( Ω ; O ( d ) / G ) (52)
and F ~ λ ( u , v ) : = + ∞ \widetilde{\mathcal{F}}^\lambda(u,v) := +\infty F λ ( u , v ) := + ∞ otherwise in L 1 ( Ω ; M d × d / G ) × L 1 ( Ω ) L^1(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}) \times L^1(\Omega) L 1 ( Ω ; M d × d / G ) × L 1 ( Ω ) where, for any u ∈ SBV 0 ( Ω ; O ( d ) / G ) u \in\operatorname{SBV}_0(\Omega;\mathrm{O}(d)/\mathcal{G}) u ∈ SBV 0 ( Ω ; O ( d ) / G ) , β ∈ SBV 0 ( Ω ; O ( d ) ) \beta\in\operatorname{SBV}_0(\Omega;\mathrm{O}(d)) β ∈ SBV 0 ( Ω ; O ( d )) is a lifting of u u u as in Theorem 2.12. The invariance (1.14) guarantees that the result does not depend on the choice of the lifting.
In this general framework we can now state the counterpart of Theorem 1.1.
Theorem 2.15. The following properties hold:
(C)’ For any ( u ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d / G × [ 0 , 1 ] ) (u_\varepsilon,v_\varepsilon)_\varepsilon\subset L^1(\Omega;\mathbb{M}^{d\times d}/\mathcal{G} \times[0,1]) ( u ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d / G × [ 0 , 1 ]) satisfying
lim sup ε → 0 F ~ ε ( u ε , v ε ) < + ∞ (53) \limsup_{\varepsilon\to0} \widetilde{\mathcal{F}}_\varepsilon(u_\varepsilon,v_\varepsilon) < +\infty \tag*{(53)} ε → 0 lim sup F ε ( u ε , v ε ) < + ∞ (53) there exists u ∈ SBV 0 ( Ω ; O ( d ) / G ) u \in\operatorname{SBV}_0(\Omega;\mathrm{O}(d)/\mathcal{G}) u ∈ SBV 0 ( Ω ; O ( d ) / G ) such that, up to a subsequence,
u ε → u in L 1 ( Ω ; M d × d / G ) , v ε → 1 in L 1 ( Ω ) . (54) u_\varepsilon\to u \text{ in } L^1(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}), \qquad v_\varepsilon\to1 \text{ in } L^1(\Omega). \tag*{(54)} u ε → u in L 1 ( Ω ; M d × d / G ) , v ε → 1 in L 1 ( Ω ) . (54) (LB)’ For any ( u ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d / G × [ 0 , 1 ] ) (u_\varepsilon,v_\varepsilon)_\varepsilon\subset L^1(\Omega;\mathbb{M}^{d\times d}/\mathcal{G} \times[0,1]) ( u ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d / G × [ 0 , 1 ]) and u ∈ SBV 0 ( Ω ; O ( d ) / G ) u \in\operatorname{SBV}_0(\Omega;\mathrm{O}(d)/\mathcal{G}) u ∈ SBV 0 ( Ω ; O ( d ) / G ) satisfying (2.34), it holds
F ~ λ ( u , 1 ) ≤ lim inf ε → 0 F ~ ε ( u ε , v ε ) . \widetilde{\mathcal{F}}^\lambda(u,1) \leq\liminf_{\varepsilon\to0} \widetilde{\mathcal{F}}_\varepsilon(u_\varepsilon,v_\varepsilon). F λ ( u , 1 ) ≤ ε → 0 lim inf F ε ( u ε , v ε ) .
(UB)’ For any u ∈ S B V 0 ( Ω ; O ( d ) / G ) u \in\mathrm{SBV}_{0}(\Omega; \mathrm{O}(d)/G) u ∈ SBV 0 ( Ω ; O ( d ) / G ) there exists ( u ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d / G ) × L 1 ( Ω ; [ 0 , 1 ] ) (u_{\varepsilon},v_{\varepsilon})_{\varepsilon} \subset L^{1}(\Omega; \mathbb{M}^{d \times d}/G) \times L^{1}(\Omega;[0,1]) ( u ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d / G ) × L 1 ( Ω ; [ 0 , 1 ]) satisfying (2.34) and
F ~ λ ( u , 1 ) ≥ lim sup ε → 0 F ~ ε ( u ε , v ε ) . (55) \widetilde{\mathcal{F}}^{\lambda}(u,1) \ge\limsup_{\varepsilon\to0} \widetilde{\mathcal{F}}_{\varepsilon}(u_{\varepsilon},v_{\varepsilon}). \tag*{(55)} F λ ( u , 1 ) ≥ ε → 0 lim sup F ε ( u ε , v ε ) . (55)
The proof of this result is obtained as a consequence of the analysis performed in the rest of the paper. Compactness (C’) is stated and proved in Theorem 3.2. The lower bound (LB’) is stated and proved in Theorem 5.1. The upper bound is a consequence of the corresponding upper bound in Theorem 1.1, proved in Section 5, Theorem 5.6, and the fact that F ~ ε ≤ F ε \widetilde{\mathcal{F}}_{\varepsilon} \le F_{\varepsilon} F ε ≤ F ε as in (2.31).
Compactness
We start with a truncation argument, in order to restrict our attention to sequences equibounded in L ∞ L^{\infty} L ∞ . We denote by Φ ( ξ ) \Phi(\xi) Φ ( ξ ) the orthogonal projection of a matrix ξ ∈ M d × d \xi\in\mathbb{M}^{d \times d} ξ ∈ M d × d on the closed ball B ‾ d : = { η ∈ M d × d : ∣ η ∣ ≤ d } \overline{B}_{\sqrt{d}} := \{\eta\in\mathbb{M}^{d \times d}: |\eta| \le\sqrt{d}\} B d := { η ∈ M d × d : ∣ η ∣ ≤ d } ,
Φ ( ξ ) : = { ξ if ∣ ξ ∣ ≤ d , d ξ ∣ ξ ∣ if ∣ ξ ∣ > d . (56) \Phi(\xi) :=
\begin{cases}
\xi& \text{if } |\xi| \le\sqrt{d},\\
\sqrt{d}\dfrac{\xi}{|\xi|} & \text{if } |\xi| > \sqrt{d}.
\end{cases} \tag*{(56)} Φ ( ξ ) := ⎩ ⎨ ⎧ ξ d ∣ ξ ∣ ξ if ∣ ξ ∣ ≤ d , if ∣ ξ ∣ > d . (56)
Since Φ ( G ξ ) = G Φ ( ξ ) \Phi(G\xi) = G\Phi(\xi) Φ ( G ξ ) = G Φ ( ξ ) for all G ∈ O ( d ) G \in\mathrm{O}(d) G ∈ O ( d ) , Φ \Phi Φ can be identified with a map from M d × d / G \mathbb{M}^{d \times d}/G M d × d / G to itself.
Lemma 3.1. Let ( u ε ) ε ⊂ H 1 ( Ω ; M d × d / G ) (u_{\varepsilon})_{\varepsilon} \subset H^{1}(\Omega;\mathbb{M}^{d \times d}/G) ( u ε ) ε ⊂ H 1 ( Ω ; M d × d / G ) be such that
lim ε → 0 ∫ Ω d G 2 ( u ε , O ( d ) / G ) d x = 0. (57) \lim_{\varepsilon\to0}\int_{\Omega} d_{G}^{2}(u_{\varepsilon},\mathrm{O}(d)/G)\,\mathrm{d}x = 0. \tag*{(57)} ε → 0 lim ∫ Ω d G 2 ( u ε , O ( d ) / G ) d x = 0. (57) Then, for every ( v ε ) ε ⊂ L 1 ( Ω ; [ 0 , 1 ] ) (v_{\varepsilon})_{\varepsilon} \subset L^{1}(\Omega;[0,1]) ( v ε ) ε ⊂ L 1 ( Ω ; [ 0 , 1 ]) we have that
F ~ ε ( Φ ( u ε ) , v ε ) ≤ F ~ ε ( u ε , v ε ) , lim ε → 0 ∥ Φ ( u ε ) − u ε ∥ L 1 ( Ω ; M d × d / G ) = 0. (58) \widetilde{\mathcal{F}}_{\varepsilon}(\Phi(u_{\varepsilon}),v_{\varepsilon}) \le\widetilde{\mathcal{F}}_{\varepsilon}(u_{\varepsilon},v_{\varepsilon}), \qquad\lim_{\varepsilon\to0}\|\Phi(u_{\varepsilon})-u_{\varepsilon}\|_{L^{1}(\Omega;\mathbb{M}^{d \times d}/G)}=0. \tag*{(58)} F ε ( Φ ( u ε ) , v ε ) ≤ F ε ( u ε , v ε ) , ε → 0 lim ∥Φ ( u ε ) − u ε ∥ L 1 ( Ω ; M d × d / G ) = 0. (58) A corresponding assertion holds for F ε F_{\varepsilon} F ε and any ( β ε ) ⊂ S B V G ( Ω ; M d × d ) (\beta_{\varepsilon}) \subset\mathrm{SBV}_{G}(\Omega;\mathbb{M}^{d \times d}) ( β ε ) ⊂ SBV G ( Ω ; M d × d ) .
Proof. It follows immediately from the definition of Φ \Phi Φ , and the fact that dist ( Φ ( ξ ) , O ( d ) ) ≤ dist ( ξ , O ( d ) ) \operatorname{dist}(\Phi(\xi),\mathrm{O}(d)) \le\operatorname{dist}(\xi,\mathrm{O}(d)) dist ( Φ ( ξ ) , O ( d )) ≤ dist ( ξ , O ( d )) for any matrix ξ \xi ξ . Moreover, Φ \Phi Φ is 1 1 1 -Lipschitz, and hence, by Definition 2.2 we deduce that ∣ ∇ Φ ( u ε ) ∣ ≤ ∣ ∇ u ε ∣ |\nabla\Phi(u_{\varepsilon})| \le|\nabla u_{\varepsilon}| ∣∇Φ ( u ε ) ∣ ≤ ∣∇ u ε ∣ . Finally, since O ( d ) ⊂ B ‾ d \mathrm{O}(d) \subset\overline{B}_{\sqrt{d}} O ( d ) ⊂ B d and Φ \Phi Φ is a projection,
∣ Φ ( ξ ) − ξ ∣ ≤ dist ( ξ , B ‾ d ) ≤ dist ( ξ , O ( d ) ) . |\Phi(\xi)-\xi| \le\operatorname{dist}(\xi,\overline{B}_{\sqrt{d}}) \le\operatorname{dist}(\xi,\mathrm{O}(d)). ∣Φ ( ξ ) − ξ ∣ ≤ dist ( ξ , B d ) ≤ dist ( ξ , O ( d )) .
We first prove compactness in Theorem 2.15 by using the theory of BV functions valued in metric spaces and the strategy from [21 ] .
Theorem 3.2. For any ( u ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d / G × [ 0 , 1 ] ) (u_{\varepsilon},v_{\varepsilon})_{\varepsilon} \subset L^{1}(\Omega;\mathbb{M}^{d \times d}/G \times[0,1]) ( u ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d / G × [ 0 , 1 ]) satisfying (2.33) there exists u ∈ S B V 0 ( Ω ; O ( d ) / G ) u \in\mathrm{SBV}_{0}(\Omega;\mathrm{O}(d)/G) u ∈ SBV 0 ( Ω ; O ( d ) / G ) such that, up to a subsequence,
u ε → u in L 1 ( Ω ; M d × d / G ) , v ε → 1 in L 1 ( Ω ) . u_{\varepsilon} \to u \text{ in } L^{1}(\Omega;\mathbb{M}^{d \times d}/G), \qquad v_{\varepsilon} \to1 \text{ in } L^{1}(\Omega). u ε → u in L 1 ( Ω ; M d × d / G ) , v ε → 1 in L 1 ( Ω ) . In particular, (C’) in Theorem 2.15 is true.
Proof. Let us fix a vanishing sequence ( ε n ) n (\varepsilon_{n})_{n} ( ε n ) n and denote F ~ n ≡ F ~ ε n \widetilde{\mathcal{F}}_{n} \equiv\widetilde{\mathcal{F}}_{\varepsilon_{n}} F n ≡ F ε n for every n ∈ N n \in\mathbb{N} n ∈ N . Let ( u n , v n ) n ⊂ L 1 ( Ω ; M d × d / G × [ 0 , 1 ] ) (u_{n},v_{n})_{n} \subset L^{1}(\Omega;\mathbb{M}^{d \times d}/G \times[0,1]) ( u n , v n ) n ⊂ L 1 ( Ω ; M d × d / G × [ 0 , 1 ]) be such that
sup n F ~ n ( u n , v n ) ≤ K . (59) \sup_{n}\widetilde{\mathcal{F}}_{n}(u_{n},v_{n}) \le K. \tag*{(59)} n sup F n ( u n , v n ) ≤ K . (59)
The convergence of v n v_{n} v n to 1 1 1 directly follows from the estimate
∫ Ω ( 1 − v n ) 2 4 ε n d x ≤ K . (60) \int_{\Omega}\frac{(1-v_{n})^{2}}{4\varepsilon_{n}}\,\mathrm{d}x \le K. \tag*{(60)} ∫ Ω 4 ε n ( 1 − v n ) 2 d x ≤ K . (60)
Moreover, u n ∈ H 1 ( Ω ; M d × d / G ) u_n \in H^1(\Omega; \mathrm{M}^{d \times d}/G) u n ∈ H 1 ( Ω ; M d × d / G ) by (59) , and, by Lemma 3.1, we may assume that
∥ u n ∥ L ∞ ≤ d for every n ∈ N . (61) \lVert u_n\rVert_{L^\infty} \le\sqrt{d} \qquad\text{for every } n \in\mathbb{N}. \tag*{(61)} ∥ u n ∥ L ∞ ≤ d for every n ∈ N . (61)
We follow the strategy from [21 Prop. 4.11] . As a first step, we need to identify small subsets of Ω \Omega Ω , with vanishing measure and bounded perimeter, where the the energy bound cannot provide any uniform bound of ∣ ∇ u n ∣ \lvert\nabla u_n\rvert ∣ ∇ u n ∣ . More precisely, fixed ρ ∈ ( 0 , 1 / 2 ) \rho\in(0,1/2) ρ ∈ ( 0 , 1/2 ) , for every n n n we can choose t n ρ ∈ ( 1 − ρ , 1 − ρ / 2 ) t_n^\rho\in(1-\rho,1-\rho/2) t n ρ ∈ ( 1 − ρ , 1 − ρ /2 ) such that
E n ρ : = { v n < λ n ρ } obeys Per ( E n ρ , Ω ) ≤ C ρ K and L N ( E n ρ ) ≤ C ρ ε n K . (62) E_n^\rho:= \{v_n < \lambda_n^\rho\} \text{ obeys } \operatorname{Per}(E_n^\rho,\Omega) \le C_\rho K \quad\text{and} \quad\mathcal{L}^N(E_n^\rho) \le C_\rho\varepsilon_n K. \tag*{(62)} E n ρ := { v n < λ n ρ } obeys Per ( E n ρ , Ω ) ≤ C ρ K and L N ( E n ρ ) ≤ C ρ ε n K . (62)
Namely, for ϕ ( t ) : = ( 1 − t ) 2 / 2 \phi(t):=(1-t)^2/2 ϕ ( t ) := ( 1 − t ) 2 /2 , by the coarea formula, Young’s inequality and (59) , we get
∫ R Per ( { ϕ ( v n ) > t } , Ω ) d t = ∫ Ω ∣ ∇ ( ϕ ( v n ) ) ∣ d x ≤ ∫ Ω ( ( 1 − v n ) 2 4 ε n + ε n ∣ ∇ v n ∣ 2 ) d x ≤ K , (63) \int_{\mathbb{R}} \operatorname{Per}(\{\phi(v_n)>t\},\Omega)\,dt=\int_\Omega\lvert\nabla(\phi(v_n))\rvert\,dx\le\int_\Omega\left(\frac{(1-v_n)^2}{4\varepsilon_n}+\varepsilon_n\lvert\nabla v_n\rvert^2\right)\,dx\le K, \tag*{(63)} ∫ R Per ({ ϕ ( v n ) > t } , Ω ) d t = ∫ Ω ∣ ∇ ( ϕ ( v n ))∣ d x ≤ ∫ Ω ( 4 ε n ( 1 − v n ) 2 + ε n ∣ ∇ v n ∣ 2 ) d x ≤ K , (63)
and it is enough to choose λ n ρ = ϕ − 1 ( t n ρ ) \lambda_n^\rho=\phi^{-1}(t_n^\rho) λ n ρ = ϕ − 1 ( t n ρ ) , with t n ρ ∈ ( ϕ ( 1 − ρ / 2 ) , ϕ ( 1 − ρ ) ) t_n^\rho\in(\phi(1-\rho/2),\phi(1-\rho)) t n ρ ∈ ( ϕ ( 1 − ρ /2 ) , ϕ ( 1 − ρ )) such that
( ϕ ( 1 − ρ ) − ϕ ( 1 − ρ / 2 ) ) Per ( { ϕ ( v n ) > t n ρ } , Ω ) ≤ ∫ ϕ ( 1 − ρ / 2 ) ϕ ( 1 − ρ ) Per ( { ϕ ( v n ) > t } , Ω ) d t . \left(\phi(1-\rho)-\phi(1-\rho/2)\right)\operatorname{Per}(\{\phi(v_n)>t_n^\rho\},\Omega)\le\int_{\phi(1-\rho/2)}^{\phi(1-\rho)}\operatorname{Per}(\{\phi(v_n)>t\},\Omega)\,dt. ( ϕ ( 1 − ρ ) − ϕ ( 1 − ρ /2 ) ) Per ({ ϕ ( v n ) > t n ρ } , Ω ) ≤ ∫ ϕ ( 1 − ρ /2 ) ϕ ( 1 − ρ ) Per ({ ϕ ( v n ) > t } , Ω ) d t .
The second estimate in (62) then follows from the fact that
L N ( { v n < λ n } ) ( 1 − λ n ) 2 4 ε n ≤ ∫ { v n < λ n } ( 1 − v n ) 2 4 ε n d x ≤ K . \mathcal{L}^N(\{v_n<\lambda_n\})\frac{(1-\lambda_n)^2}{4\varepsilon_n}\le\int_{\{v_n<\lambda_n\}}\frac{(1-v_n)^2}{4\varepsilon_n}\,dx\le K. L N ({ v n < λ n }) 4 ε n ( 1 − λ n ) 2 ≤ ∫ { v n < λ n } 4 ε n ( 1 − v n ) 2 d x ≤ K .
Now we use again (59) to get a uniform estimate for u n u_n u n in Ω ∖ E n ρ \Omega\setminus E_n^\rho Ω ∖ E n ρ , exploiting the fact that v n ≥ 1 − ρ v_n\ge1-\rho v n ≥ 1 − ρ in this set. Namely, setting B n : = { f ( v n ) > M ε n ε n − 1 / 2 } B_n:=\{f(v_n)>M\varepsilon_n\varepsilon_n^{-1/2}\} B n := { f ( v n ) > M ε n ε n − 1/2 } , we have
K ≥ F ‾ n ( u n , v n ; Ω ∖ E n ρ ) ≥ M ε n 2 ∫ B n ∖ E n ρ ∣ ∇ u n ∣ 2 d x + ∫ Ω ∖ B n ∖ E n ρ ( ε n f 2 ( v n ) ∣ ∇ u n ∣ 2 + ( 1 − v n ) 2 4 ε n ) d x ≥ M ε n 2 ∫ B n ∖ E n ρ ∣ ∇ u n ∣ 2 d x + ∫ Ω ∖ B n ∖ E n ρ f ( v n ) ( 1 − v n ) ∣ ∇ u n ∣ d x , \begin{aligned}
K&\ge\overline{F}_n(u_n,v_n;\Omega\setminus E_n^\rho)\ge M_{\varepsilon_n}^2\int_{B_n\setminus E_n^\rho}\lvert\nabla u_n\rvert^2\,dx\\
&\quad+\int_{\Omega\setminus B_n\setminus E_n^\rho}\left(\varepsilon_n f^2(v_n)\lvert\nabla u_n\rvert^2+\frac{(1-v_n)^2}{4\varepsilon_n}\right)\,dx\\
&\ge M_{\varepsilon_n}^2\int_{B_n\setminus E_n^\rho}\lvert\nabla u_n\rvert^2\,dx+\int_{\Omega\setminus B_n\setminus E_n^\rho}f(v_n)(1-v_n)\lvert\nabla u_n\rvert\,dx,
\end{aligned} K ≥ F n ( u n , v n ; Ω ∖ E n ρ ) ≥ M ε n 2 ∫ B n ∖ E n ρ ∣ ∇ u n ∣ 2 d x + ∫ Ω ∖ B n ∖ E n ρ ( ε n f 2 ( v n ) ∣ ∇ u n ∣ 2 + 4 ε n ( 1 − v n ) 2 ) d x ≥ M ε n 2 ∫ B n ∖ E n ρ ∣ ∇ u n ∣ 2 d x + ∫ Ω ∖ B n ∖ E n ρ f ( v n ) ( 1 − v n ) ∣ ∇ u n ∣ d x ,
so that
∫ Ω ∖ E n ρ ∣ ∇ u n ∣ d x ≤ ( K L N ( Ω ) ) 1 / 2 M ε n + K m ρ , (64) \int_{\Omega\setminus E_n^\rho}\lvert\nabla u_n\rvert\,dx\le\frac{(K\mathcal{L}^N(\Omega))^{1/2}}{M_{\varepsilon_n}}+\frac{K}{m_\rho}, \tag*{(64)} ∫ Ω ∖ E n ρ ∣ ∇ u n ∣ d x ≤ M ε n ( K L N ( Ω ) ) 1/2 + m ρ K , (64)
where 0 < m ρ : = min s ∈ [ 1 − ρ , 1 ) ( 1 − s ) f ( s ) 0<m_\rho:=\min_{s\in[1-\rho,1)}(1-s)f(s) 0 < m ρ := min s ∈ [ 1 − ρ , 1 ) ( 1 − s ) f ( s ) obeys lim ρ → 0 m ρ = ∞ \lim_{\rho\to0}m_\rho=\infty lim ρ → 0 m ρ = ∞ . In conclusion, setting
u ‾ n ρ : = u n χ Ω ∖ E n ρ + [ I d ] χ E n ρ , (65) \overline{u}_n^\rho:=u_n\chi_{\Omega\setminus E_n^\rho}+[{\rm Id}]\chi_{E_n^\rho}, \tag*{(65)} u n ρ := u n χ Ω ∖ E n ρ + [ Id ] χ E n ρ , (65)
we have u ‾ n ρ ∈ S B V ( Ω ; M d × d / G ) \overline{u}_n^\rho\in SBV(\Omega;\mathrm{M}^{d\times d}/G) u n ρ ∈ S B V ( Ω ; M d × d / G ) with
J u ‾ n ρ ⊂ ∂ ∗ E n ρ , ∣ ∇ u ‾ n ρ ∣ = ∣ ∇ u n ∣ χ Ω ∖ E n ρ = ∣ ∇ ( u n χ Ω ∖ E n ρ ) ∣ . (66) J_{\overline{u}_n^\rho}\subset\partial^*E_n^\rho,\qquad\lvert\nabla\overline{u}_n^\rho\rvert=\lvert\nabla u_n\rvert\chi_{\Omega\setminus E_n^\rho}=\lvert\nabla(u_n\chi_{\Omega\setminus E_n^\rho})\rvert. \tag*{(66)} J u n ρ ⊂ ∂ ∗ E n ρ , ∣ ∇ u n ρ ∣ = ∣ ∇ u n ∣ χ Ω ∖ E n ρ = ∣ ∇ ( u n χ Ω ∖ E n ρ )∣ . (66)
By (61) , (62) , (66) , (64) and Theorem 2.10 we obtain that there exist a function u ρ ∈ B V ( Ω ; M d × d / G ) u^\rho\in BV(\Omega;\mathrm{M}^{d\times d}/G) u ρ ∈ B V ( Ω ; M d × d / G ) and a subsequence ( u ‾ n k ρ ) k (\overline{u}_{n_k}^\rho)_k ( u n k ρ ) k converging to u ρ u^\rho u ρ in L 1 ( Ω ; M d × d / G ) L^1(\Omega;\mathrm{M}^{d\times d}/G) L 1 ( Ω ; M d × d / G ) . Since (59) implies (57) , we obtain that u ρ u^\rho u ρ takes values in O ( d ) / G O(d)/G O ( d ) / G almost everywhere.
Moreover, since L N ( E n ρ ) ≤ C ρ ε n K → 0 \mathcal{L}^N(E_n^\rho)\le C_\rho\varepsilon_nK\to0 L N ( E n ρ ) ≤ C ρ ε n K → 0 and ∥ u n ∥ L ∞ ≤ d \lVert u_n\rVert_{L^\infty}\le\sqrt d ∥ u n ∥ L ∞ ≤ d , we have ∥ u n − u ‾ n ρ ∥ L 1 → 0 \lVert u_n-\overline{u}_n^\rho\rVert_{L^1}\to0 ∥ u n − u n ρ ∥ L 1 → 0 , therefore also the subsequence ( u n k ) k (u_{n_k})_k ( u n k ) k converges to the same u ρ u^\rho u ρ . In particular, taking a diagonal subsequence we can ensure that the limit does not depend on ρ \rho ρ , we denote it by u u u .
Let φ ∈ Lip 1 ( M d × d / G ) \varphi\in\operatorname{Lip}_1(\mathrm{M}^{d\times d}/G) φ ∈ Lip 1 ( M d × d / G ) . Then for any ρ \rho ρ and n n n we have φ ∘ u ‾ n ρ ∈ B V ( Ω ) \varphi\circ\overline{u}_n^\rho\in BV(\Omega) φ ∘ u n ρ ∈ B V ( Ω ) ; from (62) and (66) we obtain sup n H N − 1 ( J φ ∘ u ‾ n ρ ) < ∞ \sup_n\mathcal{H}^{N-1}(J_{\varphi\circ\overline{u}_n^\rho})<\infty sup n H N − 1 ( J φ ∘ u n ρ ) < ∞ . Therefore, recalling (64) , and exploiting a standard localization argument, e.g. arguing by slicing to reduce to a 1-dimensional domain where the approximating functions have the same (finite) number of jump points, we get
∣ D c ( φ ∘ u ) ∣ ( Ω ) + ∥ ∇ ( φ ∘ u ) ∥ L 1 ( Ω ) ≤ lim inf k → ∞ ∥ ∇ ( φ ∘ u ‾ n k ) ∥ L 1 ( Ω ) ≤ lim inf k → ∞ ∥ ∇ u n k ∥ L 1 ( Ω ∖ E n k ρ ) ≤ K m ρ . (67) \left|D^{c}(\varphi\circ u)\right|(\Omega)+\|\nabla(\varphi\circ u)\|_{L^{1}(\Omega)} \le\liminf_{k\to\infty}\|\nabla(\varphi\circ\overline{u}_{n_k})\|_{L^{1}(\Omega)}
\le\liminf_{k\to\infty}\|\nabla u_{n_k}\|_{L^{1}(\Omega\setminus E^{\rho}_{n_k})}\le\frac{K}{m_{\rho}}. \tag*{(67)} ∣ D c ( φ ∘ u ) ∣ ( Ω ) + ∥∇ ( φ ∘ u ) ∥ L 1 ( Ω ) ≤ k → ∞ lim inf ∥∇ ( φ ∘ u n k ) ∥ L 1 ( Ω ) ≤ k → ∞ lim inf ∥∇ u n k ∥ L 1 ( Ω ∖ E n k ρ ) ≤ m ρ K . (67)
We recall that lim ρ → 0 m ρ = ∞ \lim_{\rho\to0}m_{\rho}=\infty lim ρ → 0 m ρ = ∞ . Since the left-hand side does not depend on ρ \rho ρ , it is zero, and we obtain φ ∘ u ∈ S B V 0 \varphi\circ u\in\mathrm{SBV}_{0} φ ∘ u ∈ SBV 0 . By Remark 2.9, u ∈ S B V 0 ( Ω ; O ( d ) / G ) u\in\mathrm{SBV}_{0}(\Omega;\mathrm{O}(d)/G) u ∈ SBV 0 ( Ω ; O ( d ) / G ) . □
We remark that the compactness in Theorem 1.1 is a direct consequence of Theorem 3.2 , owing to the properties in Remark 2.4 and Theorem 2.12 on the lifting of BV functions with values in O ( d ) / G \mathrm{O}(d)/G O ( d ) / G , applied to any limiting field u u u .
Nevertheless, we provide a different proof of Theorem 1.1 which does not make use of the theory of BV functions valued in metric spaces. We exploit the fact that we may apply Theorem 2.12 to functions valued in O ( d ) / G \mathrm{O}(d)/G O ( d ) / G , target set of the limits. The crucial step is to modify the original sequence in such a way that the modifications are valued into O ( d ) / G \mathrm{O}(d)/G O ( d ) / G and still have uniformly bounded energy. By this approach we prove weak* convergence in BV, up to a small perturbation.
Theorem 3.3. Let ( β ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d × [ 0 , 1 ] ) (\beta_{\varepsilon},v_{\varepsilon})_{\varepsilon}\subset L^{1}(\Omega;\mathbb{M}^{d\times d}\times[0,1]) ( β ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d × [ 0 , 1 ]) satisfy (1.15). Then there exists β ∈ S B V 0 ( Ω ; O ( d ) ) \beta\in\mathrm{SBV}_{0}(\Omega;\mathrm{O}(d)) β ∈ SBV 0 ( Ω ; O ( d )) , a subsequence ( ε n ) n (\varepsilon_n)_n ( ε n ) n and G n ∈ L ∞ ( Ω ; G ) G_n\in L^{\infty}(\Omega;G) G n ∈ L ∞ ( Ω ; G ) such that G n β ε n G_n\beta_{\varepsilon_n} G n β ε n belong to S B V G ( Ω ; M d × d ) \mathrm{SBV}_{G}(\Omega;\mathbb{M}^{d\times d}) SBV G ( Ω ; M d × d ) and converge to β \beta β in L 1 ( Ω ; M d × d ) L^{1}(\Omega;\mathbb{M}^{d\times d}) L 1 ( Ω ; M d × d ) , and v ε n → 1 v_{\varepsilon_n}\to1 v ε n → 1 in L 2 ( Ω ) L^{2}(\Omega) L 2 ( Ω ) . In particular, (C) in Theorem 1.1 is true.
Proof. By Lemma 3.1 we may assume that
∣ β ε ∣ ≤ d a.e. in Ω . (68) |\beta_{\varepsilon}|\le\sqrt{d}\quad\text{a.e. in }\Omega. \tag*{(68)} ∣ β ε ∣ ≤ d a.e. in Ω. (68)
Up to passing to a subsequence ( β ε n ) n ≡ ( β n ) n (\beta_{\varepsilon_n})_n\equiv(\beta_n)_n ( β ε n ) n ≡ ( β n ) n , by (1.15) we get
sup n F ε n ( β n , v n ) ≤ K . (69) \sup_n \mathcal{F}_{\varepsilon_n}(\beta_n,v_n)\le K. \tag*{(69)} n sup F ε n ( β n , v n ) ≤ K . (69)
We deal with ( v n ) n (v_n)_n ( v n ) n as in the proof of Theorem 3.2 ; in particular, v n → 1 v_n\to1 v n → 1 a.e. in Ω \Omega Ω and for any ρ ∈ ( 0 , 1 / 2 ) \rho\in(0,1/2) ρ ∈ ( 0 , 1/2 ) there are suitable t n ρ ∈ ( 1 − ρ , 1 − ρ / 2 ) t_n^\rho\in(1-\rho,1-\rho/2) t n ρ ∈ ( 1 − ρ , 1 − ρ /2 ) such that E n ρ : = { v n < t n ρ } E_n^\rho:=\{v_n<t_n^\rho\} E n ρ := { v n < t n ρ } satisfies for every n ∈ N n\in\mathbb{N} n ∈ N the properties (3.7), and the same computation as for (3.9) leads to
∫ Ω ∖ E n ρ ∣ ∇ β n ∣ d x ≤ ( K L N ( Ω ) ) 1 / 2 M ε n + K m ρ . (70) \int_{\Omega\setminus E_n^\rho}|\nabla\beta_n|\,\mathrm{d}x\le\frac{(K\mathcal{L}^{N}(\Omega))^{1/2}}{M_{\varepsilon_n}}+\frac{K}{m_{\rho}}. \tag*{(70)} ∫ Ω ∖ E n ρ ∣∇ β n ∣ d x ≤ M ε n ( K L N ( Ω ) ) 1/2 + m ρ K . (70)
The next part of the argument differs from the one in Theorem 3.2 . We consider the functions
h n ( x ) : = β n T ( x ) β n ( x ) , n ∈ N . (71) h_n(x):=\beta_n^{T}(x)\beta_n(x),\qquad n\in\mathbb{N}. \tag*{(71)} h n ( x ) := β n T ( x ) β n ( x ) , n ∈ N . (71)
We have that h n ∈ S B V ( Ω ; M s y m d × d ) h_n\in\mathrm{SBV}(\Omega;\mathbb{M}^{d\times d}_{\mathrm{sym}}) h n ∈ SBV ( Ω ; M sym d × d ) with ∥ h n ∥ L ∞ ≤ d \|h_n\|_{L^\infty}\le d ∥ h n ∥ L ∞ ≤ d , M s y m d × d \mathbb{M}^{d\times d}_{\mathrm{sym}} M sym d × d denoting the set of d × d d\times d d × d symmetric matrices, and J h n ⊂ J β n J_{h_n}\subset J_{\beta_n} J h n ⊂ J β n . Moreover, if x ∈ J h n x\in J_{h_n} x ∈ J h n , and G ∈ G G\in G G ∈ G is such that β n + ( x ) = G β n − ( x ) \beta_n^{+}(x)=G\beta_n^{-}(x) β n + ( x ) = G β n − ( x ) , then
h n + ( x ) = ( β n + ( x ) ) T β n + ( x ) = ( G β n − ( x ) ) T G β n − ( x ) = ( β n − ( x ) ) T β n − ( x ) = h n − ( x ) , h_n^{+}(x)=(\beta_n^{+}(x))^{T}\beta_n^{+}(x)=(G\beta_n^{-}(x))^{T}G\beta_n^{-}(x)=(\beta_n^{-}(x))^{T}\beta_n^{-}(x)=h_n^{-}(x), h n + ( x ) = ( β n + ( x ) ) T β n + ( x ) = ( G β n − ( x ) ) T G β n − ( x ) = ( β n − ( x ) ) T β n − ( x ) = h n − ( x ) ,
therefore h n ∈ W 1 , 1 ( Ω ; M s y m d × d ) h_n\in W^{1,1}(\Omega;\mathbb{M}^{d\times d}_{\mathrm{sym}}) h n ∈ W 1 , 1 ( Ω ; M sym d × d ) , with ∣ ∇ h n ∣ ≤ C ∣ ∇ β n ∣ |\nabla h_n|\le C|\nabla\beta_n| ∣∇ h n ∣ ≤ C ∣∇ β n ∣ a.e. in Ω \Omega Ω .
A similar computation shows that
w n : = ∣ det β n ∣ ∈ W 1 , 1 ( Ω ) , w_n:=|\det\beta_n|\in W^{1,1}(\Omega), w n := ∣ det β n ∣ ∈ W 1 , 1 ( Ω ) ,
with ∣ ∇ w n ∣ ≤ C ∣ ∇ β n ∣ |\nabla w_n|\le C|\nabla\beta_n| ∣∇ w n ∣ ≤ C ∣∇ β n ∣ a.e. in Ω \Omega Ω . By the Coarea Formula, for every n ∈ N n\in\mathbb{N} n ∈ N there exists λ n ′ ∈ ( 1 4 , 1 2 ) \lambda_n'\in\left(\frac14,\frac12\right) λ n ′ ∈ ( 4 1 , 2 1 ) such that
Per ( { ∣ det β n ∣ < λ n ′ } , Ω ) ≤ C . (72) \operatorname{Per}(\{|\det\beta_n|<\lambda_n'\},\Omega)\le C. \tag*{(72)} Per ({ ∣ det β n ∣ < λ n ′ } , Ω ) ≤ C . (72)
Moreover, since λ n ≤ 1 / 2 \lambda_n \le1/2 λ n ≤ 1/2 , and min { dist ( A , O ( d ) ) : ∣ det A ∣ ≤ 1 / 2 } > 0 \min\{\operatorname{dist}(A,O(d)) : |\det A| \le1/2\} > 0 min { dist ( A , O ( d )) : ∣ det A ∣ ≤ 1/2 } > 0 ,
L N ( { ∣ det β n ∣ < λ n ′ } ) ≤ C ∫ Ω dist 2 ( β n , O ( d ) ) d x ≤ C K δ ε n . \mathcal{L}^{N}(\{|\det\beta_n| < \lambda'_n\}) \le C \int_{\Omega} \operatorname{dist}^{2}(\beta_n,O(d))\,dx \le CK\delta\varepsilon_n. L N ({ ∣ det β n ∣ < λ n ′ }) ≤ C ∫ Ω dist 2 ( β n , O ( d )) d x ≤ C K δ ε n .
Hence, the sets F n ρ : = E n ρ ∪ { ∣ det β n ∣ < λ n ′ } F_n^\rho:= E_n^\rho\cup\{|\det\beta_n| < \lambda'_n\} F n ρ := E n ρ ∪ { ∣ det β n ∣ < λ n ′ } obey
Per ( F n ρ , Ω ) ≤ C + C ρ K and L N ( F n ρ ) ≤ C ρ ε n K + C K δ ε n . (73) \operatorname{Per}(F_n^\rho,\Omega) \le C+C_\rho K \quad\text{and}\quad\mathcal{L}^{N}(F_n^\rho) \le C_\rho\varepsilon_n K+CK\delta\varepsilon_n. \tag*{(73)} Per ( F n ρ , Ω ) ≤ C + C ρ K and L N ( F n ρ ) ≤ C ρ ε n K + C K δ ε n . (73)
and the functions
h ˉ n : = h n χ Ω ∖ F n ρ + Id χ F n ρ (74) \bar h_n := h_n\chi_{\Omega\setminus F_n^\rho}+\operatorname{Id}\chi_{F_n^\rho} \tag*{(74)} h ˉ n := h n χ Ω ∖ F n ρ + Id χ F n ρ (74)
satisfy h ˉ n ∈ SBV ( Ω ; M s y m d × d ) \bar h_n\in\operatorname{SBV}(\Omega;\mathbb{M}^{d\times d}_{\mathrm{sym}}) h ˉ n ∈ SBV ( Ω ; M sym d × d ) with
∣ D h ˉ n ∣ ( Ω ) + ∥ h ˉ n ∥ L ∞ ( Ω ; M d × d ) ≤ C , det ( h ˉ n ) ≥ 1 16 a.e. in Ω . (75) |D\bar h_n|(\Omega)+\|\bar h_n\|_{L^\infty(\Omega;\mathbb{M}^{d\times d})}\le C,\qquad\det(\bar h_n)\ge\frac{1}{16}\quad\text{a.e. in }\Omega. \tag*{(75)} ∣ D h ˉ n ∣ ( Ω ) + ∥ h ˉ n ∥ L ∞ ( Ω ; M d × d ) ≤ C , det ( h ˉ n ) ≥ 16 1 a.e. in Ω. (75)
The maps A ↦ A 2 A\mapsto A^2 A ↦ A 2 and A ↦ A − 1 A\mapsto A^{-1} A ↦ A − 1 are smooth bijections of M s y m , + d × d \mathbb{M}^{d\times d}_{\mathrm{sym},+} M sym , + d × d , i.e., the set of strictly positive-definite matrices in M s y m d × d \mathbb{M}^{d\times d}_{\mathrm{sym}} M sym d × d , onto itself with smooth inverse, and in particular their inverses A ↦ A 1 / 2 A\mapsto A^{1/2} A ↦ A 1/2 and A ↦ A − 1 / 2 A\mapsto A^{-1/2} A ↦ A − 1/2 are smooth, and then Lipschitz on compact sets. Therefore the functions ( h ˉ n ) − 1 / 2 (\bar h_n)^{-1/2} ( h ˉ n ) − 1/2 , defined by ( h ˉ n ) − 1 / 2 ( x ) : = ( ( h ˉ n ( x ) − 1 ) 1 / 2 ) (\bar h_n)^{-1/2}(x):=((\bar h_n(x)^{-1})^{1/2}) ( h ˉ n ) − 1/2 ( x ) := (( h ˉ n ( x ) − 1 ) 1/2 ) , are in SBV ( Ω ; M s y m d × d ) \operatorname{SBV}(\Omega;\mathbb{M}^{d\times d}_{\mathrm{sym}}) SBV ( Ω ; M sym d × d ) , and
J ( h ˉ n ) − 1 / 2 = J h ˉ n ⊂ ∂ ∗ F n ρ , ∣ ∇ ( ( h ˉ n ) 1 / 2 ) ∣ + ∣ ∇ ( ( h ˉ n ) − 1 / 2 ) ∣ ≤ C ∣ ∇ h ˉ n ∣ a.e. in Ω ∖ F n ρ . (76) J_{(\bar h_n)^{-1/2}}=J_{\bar h_n}\subset\partial^*F_n^\rho,\qquad|\nabla((\bar h_n)^{1/2})|+|\nabla((\bar h_n)^{-1/2})|\le C|\nabla\bar h_n|\quad\text{a.e. in }\Omega\setminus F_n^\rho. \tag*{(76)} J ( h ˉ n ) − 1/2 = J h ˉ n ⊂ ∂ ∗ F n ρ , ∣∇ (( h ˉ n ) 1/2 ) ∣ + ∣∇ (( h ˉ n ) − 1/2 ) ∣ ≤ C ∣∇ h ˉ n ∣ a.e. in Ω ∖ F n ρ . (76)
In particular, the functions k n : Ω → O ( d ) k_n:\Omega\to O(d) k n : Ω → O ( d ) defined by
k n : = β n ( h ˉ n ) − 1 / 2 χ Ω ∖ F n ρ + Id χ F n ρ , (77) k_n:=\beta_n(\bar h_n)^{-1/2}\chi_{\Omega\setminus F_n^\rho}+\operatorname{Id}\chi_{F_n^\rho}, \tag*{(77)} k n := β n ( h ˉ n ) − 1/2 χ Ω ∖ F n ρ + Id χ F n ρ , (77)
by (3.20), (3.21) satisfy
∣ ∇ k n ∣ ≤ C ∣ ∇ β n ∣ a.e. in Ω . |\nabla k_n|\le C|\nabla\beta_n|\quad\text{a.e. in }\Omega. ∣∇ k n ∣ ≤ C ∣∇ β n ∣ a.e. in Ω.
From (3.15) and since k n = Id k_n=\operatorname{Id} k n = Id in F n ρ F_n^\rho F n ρ (so in particular in E n ρ ⊂ F n ρ E_n^\rho\subset F_n^\rho E n ρ ⊂ F n ρ ), we deduce that k n ∈ SBV ( Ω ; O ( d ) ) k_n\in\operatorname{SBV}(\Omega;O(d)) k n ∈ SBV ( Ω ; O ( d )) with
∫ Ω ∣ ∇ k n ∣ d x ≤ C K 1 / 2 M ε n + C K m ρ , J k n ⊂ J β n ∪ ∂ ∗ F n ρ , k n + ∈ G k n − on J β n ∖ F n ρ . (78) \int_{\Omega}|\nabla k_n|\,dx\le\frac{CK^{1/2}}{M\varepsilon_n}+\frac{CK}{m_\rho},\qquad J_{k_n}\subset J_{\beta_n}\cup\partial^*F_n^\rho,\qquad k_n^+\in Gk_n^-\text{ on }J_{\beta_n}\setminus F_n^\rho. \tag*{(78)} ∫ Ω ∣∇ k n ∣ d x ≤ M ε n C K 1/2 + m ρ C K , J k n ⊂ J β n ∪ ∂ ∗ F n ρ , k n + ∈ G k n − on J β n ∖ F n ρ . (78)
Passing to the equivalence classes on the image, the functions defined by [ k n ] ( x ) : = [ k n ( x ) ] [k_n](x):=[k_n(x)] [ k n ] ( x ) := [ k n ( x )] for x ∈ Ω x\in\Omega x ∈ Ω belong to SBV ( Ω ; O ( d ) / G ) \operatorname{SBV}(\Omega;O(d)/G) SBV ( Ω ; O ( d ) / G ) , and
∣ D ( [ k n ] ) ∣ ( d 2 ) ( Ω ) ≤ ∫ Ω ∣ ∇ k n ∣ d x + ∥ k n ∥ ∞ Per ( F n ρ ) ≤ C . |D([k_n])|_{(d^2)}(\Omega)\le\int_{\Omega}|\nabla k_n|\,dx+\|k_n\|_{\infty}\operatorname{Per}(F_n^\rho)\le C. ∣ D ([ k n ]) ∣ ( d 2 ) ( Ω ) ≤ ∫ Ω ∣∇ k n ∣ d x + ∥ k n ∥ ∞ Per ( F n ρ ) ≤ C .
By Theorem 2.12, for every n ∈ N n\in\mathbb{N} n ∈ N we find G n ∈ L ∞ ( Ω ; G ) G_n\in L^\infty(\Omega;G) G n ∈ L ∞ ( Ω ; G ) such that G n k n ∈ SBV ( Ω ; O ( d ) ) G_nk_n\in\operatorname{SBV}(\Omega;O(d)) G n k n ∈ SBV ( Ω ; O ( d )) , and ∣ D ( G n k n ) ∣ ( Ω ) ≤ C |D(G_nk_n)|(\Omega)\le C ∣ D ( G n k n ) ∣ ( Ω ) ≤ C , so that the functions
β ^ n : = G n k n h ˉ n χ Ω ∖ F n ρ + Id χ F n ρ = G n β n χ Ω ∖ F n ρ + Id χ F n ρ (79) \widehat{\beta}_n:=G_nk_n\bar h_n\chi_{\Omega\setminus F_n^\rho}+\operatorname{Id}\chi_{F_n^\rho}=G_n\beta_n\chi_{\Omega\setminus F_n^\rho}+\operatorname{Id}\chi_{F_n^\rho} \tag*{(79)} β n := G n k n h ˉ n χ Ω ∖ F n ρ + Id χ F n ρ = G n β n χ Ω ∖ F n ρ + Id χ F n ρ (79)
belong to SBV ( Ω ; M d × d ) \operatorname{SBV}(\Omega;\mathbb{M}^{d\times d}) SBV ( Ω ; M d × d ) with det β ^ n ( x ) ≠ 0 \det\widehat{\beta}_n(x)\ne0 det β n ( x ) = 0 for a.e. x ∈ Ω x\in\Omega x ∈ Ω , and have equibounded BV norms. Hence there exists a function β ρ ∈ BV ( Ω ; M d × d ) \beta^\rho\in\operatorname{BV}(\Omega;\mathbb{M}^{d\times d}) β ρ ∈ BV ( Ω ; M d × d ) such that, up to a subsequence, ( β ^ n ) (\widehat{\beta}_n) ( β n ) converge to β ρ \beta^\rho β ρ in the weak* topology of BV ( Ω ; M d × d ) \operatorname{BV}(\Omega;\mathbb{M}^{d\times d}) BV ( Ω ; M d × d ) . Recalling (3.18), we infer that
G n β n χ Ω ∖ F n ρ + β n χ F n ρ → β ρ in L 1 ( Ω ; M d × d ) , (80) G_n\beta_n\chi_{\Omega\setminus F_n^\rho}+\beta_n\chi_{F_n^\rho}\to\beta^\rho\quad\text{in }L^1(\Omega;\mathbb{M}^{d\times d}), \tag*{(80)} G n β n χ Ω ∖ F n ρ + β n χ F n ρ → β ρ in L 1 ( Ω ; M d × d ) , (80)
so that the functions G n G_n G n as in the statement may be taken equal to Id \operatorname{Id} Id in F n F_n F n . Notice that G n β n ∈ SBV G ( Ω ; M d × d ) G_n\beta_n\in\operatorname{SBV}_G(\Omega;\mathbb{M}^{d\times d}) G n β n ∈ SBV G ( Ω ; M d × d ) by (3.24) and (3.14). Finally, since by (3.14)
1 δ ε n ∫ A dist 2 ( β n , O ( d ) ) d x ≤ K , \frac{1}{\delta\varepsilon_n}\int_A\operatorname{dist}^{2}(\beta_n,O(d))\,dx\le K, δ ε n 1 ∫ A dist 2 ( β n , O ( d )) d x ≤ K ,
we obtain that β ρ ∈ BV ( Ω ; O ( d ) ) \beta^\rho\in\operatorname{BV}(\Omega;O(d)) β ρ ∈ BV ( Ω ; O ( d )) .
As in the final part of the proof of Theorem 3.2 , we can ensure by a diagonal argument that β ρ : = β \beta^\rho:= \beta β ρ := β does not depend on ρ \rho ρ , and that β ∈ S B V 0 ( Ω ; O ( d ) ) \beta\in SBV_0(\Omega; O(d)) β ∈ S B V 0 ( Ω ; O ( d )) . We notice that one could obtain the fact that β ∈ S B V 0 ( Ω ; O ( d ) ) \beta\in SBV_0(\Omega; O(d)) β ∈ S B V 0 ( Ω ; O ( d )) also arguing directly by slicing. □ \square □
The surface energy density
The limit energy is defined on functions with pure jump part, namely in S B V 0 SBV_0 S B V 0 . In this section we give the definition of the surface energy density and the proof of its main properties.
The surface energy density is expressed by a cell formula in one dimension. It depends on a parameter λ ∈ [ 0 , + ∞ ] \lambda\in[0,+\infty] λ ∈ [ 0 , + ∞ ] determined by the asymptotic ratio between ε \varepsilon ε and δ ε \delta_\varepsilon δ ε in the definition of the approximating functionals, see (10) in Section 1. In order to prove its main properties, such as the s ∣ log s ∣ s|\log s| s ∣ log s ∣ growth for small opening (see Proposition 4.4 ), it is convenient to recast the energy density through a 1d cell formula which is invariant by time reparametrization (Proposition 4.2).
We conclude the section with a technical result, Lemma 4.5 , used for the proof of the upper bound when λ = + ∞ \lambda=+\infty λ = + ∞ .
Let I T : = ( − T , T ) I_T := (-T,T) I T := ( − T , T ) for T > 0 T>0 T > 0 ,
F T ( β , v ) : = ∫ − T T ( f 2 ( v ) ∣ β ′ ∣ 2 + ( 1 − v ) 2 4 + ∣ v ′ ∣ 2 ) d t , ( β , v ) ∈ H 1 ( I T ; M d × d × [ 0 , 1 ] ) , F^T(\beta,v) := \int_{-T}^{T} \left(f^2(v)|\beta'|^2+\frac{(1-v)^2}{4}+|v'|^2\right)\,\mathrm{d}t,\qquad(\beta,v)\in H^1(I_T; \mathbb{M}^{d\times d}\times[0,1]), F T ( β , v ) := ∫ − T T ( f 2 ( v ) ∣ β ′ ∣ 2 + 4 ( 1 − v ) 2 + ∣ v ′ ∣ 2 ) d t , ( β , v ) ∈ H 1 ( I T ; M d × d × [ 0 , 1 ]) ,
and
F λ T ( β , v ) : = F T ( β , v ) + λ ∫ − T T dist 2 ( β , O ( d ) ) d t for λ ∈ [ 0 , + ∞ ) . F_\lambda^T(\beta,v) := F^T(\beta,v)+\lambda\int_{-T}^{T}\operatorname{dist}^2(\beta,O(d))\,\mathrm{d}t\qquad\text{for }\lambda\in[0,+\infty). F λ T ( β , v ) := F T ( β , v ) + λ ∫ − T T dist 2 ( β , O ( d )) d t for λ ∈ [ 0 , + ∞ ) .
For R − , R + ∈ O ( d ) R^-,R^+\in O(d) R − , R + ∈ O ( d ) we let
U T [ R − , R + ] : = { ( β , v ) ∈ H 1 ( I T ; M d × d × [ 0 , 1 ] ) : β ( − T ) = R − , β ( T ) = R + , v ( ± T ) = 1 } (81) \mathcal{U}^T[R^-,R^+] := \{(\beta,v)\in H^1(I_T;\mathbb{M}^{d\times d}\times[0,1]):\beta(-T)=R^-,\ \beta(T)=R^+,\ v(\pm T)=1\} \tag*{(81)} U T [ R − , R + ] := {( β , v ) ∈ H 1 ( I T ; M d × d × [ 0 , 1 ]) : β ( − T ) = R − , β ( T ) = R + , v ( ± T ) = 1 } (81)
be the set of admissible paths connecting R − R^- R − to R + R^+ R + in I T I_T I T , and with fixed boundary conditions, v = 1 v=1 v = 1 , for the damage coefficient. For λ ∈ [ 0 , + ∞ ) \lambda\in[0,+\infty) λ ∈ [ 0 , + ∞ ) we set
g λ ∗ ( R − , R + ) : = lim T ↑ + ∞ inf { F λ T ( β , v ) : ( β , v ) ∈ U T [ R − , R + ] } , (82) g_\lambda^*(R^-,R^+) := \lim_{T\uparrow+\infty}\inf\{F_\lambda^T(\beta,v):(\beta,v)\in\mathcal{U}^T[R^-,R^+]\}, \tag*{(82)} g λ ∗ ( R − , R + ) := T ↑ + ∞ lim inf { F λ T ( β , v ) : ( β , v ) ∈ U T [ R − , R + ]} , (82)
and then
g ∞ ∗ ( R − , R + ) : = sup λ > 0 g λ ∗ ( R − , R + ) . (83) g_\infty^*(R^-,R^+) := \sup_{\lambda>0}g_\lambda^*(R^-,R^+). \tag*{(83)} g ∞ ∗ ( R − , R + ) := λ > 0 sup g λ ∗ ( R − , R + ) . (83)
One easily checks that the limit in (82) exists, and it equals the infimum over all T > 0 T>0 T > 0 . In order to recover the invariance under the action of the group we then define, for λ ∈ [ 0 , ∞ ] \lambda\in[0,\infty] λ ∈ [ 0 , ∞ ] ,
g λ ( R − , R + ) : = min G ∈ G g λ ∗ ( R − , G R + ) . (84) g_\lambda(R^-,R^+) := \min_{G\in\mathcal{G}}g_\lambda^*(R^-,GR^+). \tag*{(84)} g λ ( R − , R + ) := G ∈ G min g λ ∗ ( R − , G R + ) . (84)
Since the group is finite, in the definition of g ∞ g_\infty g ∞ the minimum over G \mathcal{G} G and the supremum over λ \lambda λ commute:
g ∞ ( R − , R + ) : = min G ∈ G g ∞ ∗ ( R − , G R + ) = sup λ > 0 g λ ( R − , R + ) . (85) g_\infty(R^-,R^+) := \min_{G\in\mathcal{G}}g_\infty^*(R^-,GR^+) = \sup_{\lambda>0}g_\lambda(R^-,R^+). \tag*{(85)} g ∞ ( R − , R + ) := G ∈ G min g ∞ ∗ ( R − , G R + ) = λ > 0 sup g λ ( R − , R + ) . (85)
We first prove some basic properties of g λ g_\lambda g λ .
Proposition 4.1. For every λ ∈ [ 0 , + ∞ ] \lambda\in[0,+\infty] λ ∈ [ 0 , + ∞ ] , R R R , R − R^- R − , R + ∈ O ( d ) R^+\in O(d) R + ∈ O ( d ) , G ∈ G G\in\mathcal{G} G ∈ G it holds that (i) g λ ( R , G R ) = 0 g_\lambda(R,GR)=0 g λ ( R , GR ) = 0 ; (ii) g λ ( R − , R + ) = g λ ( R − , G R + ) g_\lambda(R^-,R^+)=g_\lambda(R^-,GR^+) g λ ( R − , R + ) = g λ ( R − , G R + ) ; (iii) g λ ( R − , R + ) = g λ ( R − ( R + ) T , I d ) g_\lambda(R^-,R^+)=g_\lambda(R^-(R^+)^{T},\mathrm{Id}) g λ ( R − , R + ) = g λ ( R − ( R + ) T , Id ) ; (iv) g λ ( R − , R + ) = g λ ( R + , R − ) g_\lambda(R^-,R^+)=g_\lambda(R^+,R^-) g λ ( R − , R + ) = g λ ( R + , R − ) ; (v) g λ ( R − , R + ) ≤ g λ ( R − , R ) + g λ ( R , R + ) g_\lambda(R^-,R^+)\leq g_\lambda(R^-,R)+g_\lambda(R,R^+) g λ ( R − , R + ) ≤ g λ ( R − , R ) + g λ ( R , R + ) , and the same for g ∗ g^* g ∗ .
Proof. Properties (i) and (ii) follow immediately from (4.4) and the fact that F T ( β , 1 ) F^{T}(\beta,1) F T ( β , 1 ) vanishes if β \beta β is a constant rotation.
Assertion (iii) follows from the fact that for every ( β , v ) ∈ U T ( R − , R + ) (\beta,v)\in U^{T}(R^{-},R^{+}) ( β , v ) ∈ U T ( R − , R + ) , setting β ~ ( t ) : = β ( t ) ( R + ) T \widetilde{\beta}(t):=\beta(t)(R^{+})^{T} β ( t ) := β ( t ) ( R + ) T one has ( β ~ , v ) ∈ U T [ R − ( R + ) T , Id ] (\widetilde{\beta},v)\in U^{T}[R^{-}(R^{+})^{T},\operatorname{Id}] ( β , v ) ∈ U T [ R − ( R + ) T , Id ] and, since R + ∈ O ( d ) R^{+}\in O(d) R + ∈ O ( d ) , ∣ β ~ ′ ∣ = ∣ β ′ ∣ |\widetilde{\beta}'|=|\beta'| ∣ β ′ ∣ = ∣ β ′ ∣ , and dist ( β ~ , O ( d ) ) = dist ( β , O ( d ) ) \operatorname{dist}(\widetilde{\beta},O(d))=\operatorname{dist}(\beta,O(d)) dist ( β , O ( d )) = dist ( β , O ( d )) , so that F λ T ( β , v ) = F λ T ( β ~ , v ) F_{\lambda}^{T}(\beta,v)=F_{\lambda}^{T}(\widetilde{\beta},v) F λ T ( β , v ) = F λ T ( β , v ) .
A similar argument is valid for β ~ ( t ) : = β ( − t ) \widetilde{\beta}(t):=\beta(-t) β ( t ) := β ( − t ) , proving that (iv) holds.
To prove (v), for any T > 0 T>0 T > 0 and any ( β 1 , v 1 ) ∈ U T [ R − , R ] (\beta_{1},v_{1})\in U^{T}[R^{-},R] ( β 1 , v 1 ) ∈ U T [ R − , R ] and ( β 2 , v 2 ) ∈ U T [ R , R + ] (\beta_{2},v_{2})\in U^{T}[R,R^{+}] ( β 2 , v 2 ) ∈ U T [ R , R + ] , we define the pair
( β , v ) ( t ) : = { ( β 1 , v 1 ) ( t + T ) , if t ≤ 0 , ( β 2 , v 2 ) ( t − T ) , if t > 0. (\beta,v)(t):=
\begin{cases}
(\beta_{1},v_{1})(t+T), & \text{if }t\leq0,\\
(\beta_{2},v_{2})(t-T), & \text{if }t>0.
\end{cases} ( β , v ) ( t ) := { ( β 1 , v 1 ) ( t + T ) , ( β 2 , v 2 ) ( t − T ) , if t ≤ 0 , if t > 0.
One checks that ( β , v ) ∈ U 2 T [ R − , R + ] (\beta,v)\in U^{2T}[R^{-},R^{+}] ( β , v ) ∈ U 2 T [ R − , R + ] , and taking the infimum in ( β 1 , v 1 ) (\beta_{1},v_{1}) ( β 1 , v 1 ) and ( β 2 , v 2 ) (\beta_{2},v_{2}) ( β 2 , v 2 ) , and the limit T → ∞ T\to\infty T → ∞ , proves the assertion for g λ ∗ g_{\lambda}^{*} g λ ∗ . Taking the minimum over G G G gives the result for g λ g_{\lambda} g λ .
In order to prove the main properties of g λ g_{\lambda} g λ it is convenient to deal with a different characterization, obtained through minimization of profiles invariant by time-reparametrization. For fixed R − , R + ∈ O ( d ) R^{-},R^{+}\in O(d) R − , R + ∈ O ( d ) , we define
U [ R − , R + ] : = { ( β , v ) ∈ H 1 ( ( 0 , 1 ) ; M d × d × [ 0 , 1 ] ) : β ( 0 ) = R − , β ( 1 ) = R + , v ( 0 ) = v ( 1 ) = 1 } . (86) U[R^{-},R^{+}]:=\{(\beta,v)\in H^{1}((0,1);M^{d\times d}\times[0,1]):
\beta(0)=R^{-},\ \beta(1)=R^{+},\ v(0)=v(1)=1\}. \tag*{(86)} U [ R − , R + ] := {( β , v ) ∈ H 1 (( 0 , 1 ) ; M d × d × [ 0 , 1 ]) : β ( 0 ) = R − , β ( 1 ) = R + , v ( 0 ) = v ( 1 ) = 1 } . (86)
and
g ‾ λ ( R − , R + ) : = inf { ∫ 0 1 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d ) ) f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t : ( β , v ) ∈ U [ R − , R + ] } (87) \overline{g}_{\lambda}(R^{-},R^{+}):=\inf\left\{\int_{0}^{1}\sqrt{(1-v)^{2}+4\lambda\operatorname{dist}^{2}(\beta,O(d))}\sqrt{f^{2}(v)|\beta'|^{2}+|v'|^{2}}\,dt:
(\beta,v)\in U[R^{-},R^{+}]\right\} \tag*{(87)} g λ ( R − , R + ) := inf { ∫ 0 1 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t : ( β , v ) ∈ U [ R − , R + ] } (87)
for λ ∈ [ 0 , + ∞ ) \lambda\in[0,+\infty) λ ∈ [ 0 , + ∞ ) , and
g ‾ ∞ ( R − , R + ) : = sup λ > 0 g ‾ λ ( R − , R + ) . (88) \overline{g}_{\infty}(R^{-},R^{+}):=\sup_{\lambda>0}\overline{g}_{\lambda}(R^{-},R^{+}). \tag*{(88)} g ∞ ( R − , R + ) := λ > 0 sup g λ ( R − , R + ) . (88)
Proposition 4.2. For every λ ∈ [ 0 , + ∞ ] \lambda\in[0,+\infty] λ ∈ [ 0 , + ∞ ] , it holds that g λ ∗ = g ‾ λ g_{\lambda}^{*}=\overline{g}_{\lambda} g λ ∗ = g λ .
Proof. The proof follows the same lines of the one of [19 Proposition 4.3] . We argue for λ ∈ [ 0 , + ∞ ) \lambda\in[0,+\infty) λ ∈ [ 0 , + ∞ ) , the case λ = + ∞ \lambda=+\infty λ = + ∞ being an easy consequence using the definitions.
The fact that g λ ∗ ≥ g ‾ λ g_{\lambda}^{*}\geq\overline{g}_{\lambda} g λ ∗ ≥ g λ follows by a direct application of Young’s inequality and the 1 1 1 -homogeneity in the derivatives of the integral functional in (4.7). Namely, for every ( β , v ) ∈ U T [ R − , R + ] (\beta,v)\in U^{T}[R^{-},R^{+}] ( β , v ) ∈ U T [ R − , R + ] we have
F λ T ( β , v ) ≥ ∫ − T T ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d ) ) f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t = ∫ 0 1 ( 1 − v ~ ) 2 + 4 λ dist 2 ( β ~ , O ( d ) ) f 2 ( v ~ ) ∣ β ~ ′ ∣ 2 + ∣ v ~ ′ ∣ 2 d t ≥ g ‾ λ ( R − , R + ) \begin{aligned}
F_{\lambda}^{T}(\beta,v)\geq\int_{-T}^{T}\sqrt{(1-v)^{2}+4\lambda\operatorname{dist}^{2}(\beta,O(d))}\sqrt{f^{2}(v)|\beta'|^{2}+|v'|^{2}}\,dt \\
&=\int_{0}^{1}\sqrt{(1-\widetilde{v})^{2}+4\lambda\operatorname{dist}^{2}(\widetilde{\beta},O(d))}\sqrt{f^{2}(\widetilde{v})|\widetilde{\beta}'|^{2}+|\widetilde{v}'|^{2}}\,dt\geq\overline{g}_{\lambda}(R^{-},R^{+})
\end{aligned} F λ T ( β , v ) ≥ ∫ − T T ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t = ∫ 0 1 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t ≥ g λ ( R − , R + )
for β ~ ( t ) = β ( 2 T t − T ) \widetilde{\beta}(t)=\beta(2Tt-T) β ( t ) = β ( 2 T t − T ) , v ~ ( t ) = v ( 2 T t − T ) \widetilde{v}(t)=v(2Tt-T) v ( t ) = v ( 2 T t − T ) , t ∈ ( 0 , 1 ) t\in(0,1) t ∈ ( 0 , 1 ) , and the inequality follows from the definition of g λ ∗ g_{\lambda}^{*} g λ ∗ .
To prove the opposite inequality, we first remark that, by an approximation argument through more regular ( β , v ) (\beta,v) ( β , v ) , for any R − , R + R^{-},R^{+} R − , R + , η ∈ ( 0 , 1 ) \eta\in(0,1) η ∈ ( 0 , 1 ) we can find competitors ( β , v ) (\beta,v) ( β , v ) belonging to W 1 , ∞ ( ( 0 , 1 ) ; M d × d × [ 0 , 1 ] ) ∩ U [ R − , R + ] W^{1,\infty}((0,1);M^{d\times d}\times[0,1])\cap U[R^{-},R^{+}] W 1 , ∞ (( 0 , 1 ) ; M d × d × [ 0 , 1 ]) ∩ U [ R − , R + ] and satisfying
∫ 0 1 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d ) ) f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t < g ‾ λ ( R − , R + ) + η . (89) \int_{0}^{1}\sqrt{(1-v)^{2}+4\lambda\operatorname{dist}^{2}(\beta,O(d))}\sqrt{f^{2}(v)|\beta'|^{2}+|v'|^{2}}\,dt<\overline{g}_{\lambda}(R^{-},R^{+})+\eta. \tag*{(89)} ∫ 0 1 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t < g λ ( R − , R + ) + η . (89)
We set
v η ( t ) : = v ( t ) ∧ ( 1 − η ) , ψ η ( t ) : = 2 ∫ 0 t η + f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d ) ) d s , t ∈ [ 0 , 1 ] , (90) v_\eta(t):=v(t)\wedge(1-\eta),\qquad\psi_\eta(t):=2\int_0^t\sqrt{\frac{\eta+f^2(v_\eta)|\beta'|^2+|v_\eta'|^2}{(1-v_\eta)^2+4\lambda\operatorname{dist}^2(\beta,\mathrm{O}(d))}}\,\mathrm{d}s,\qquad t\in[0,1], \tag*{(90)} v η ( t ) := v ( t ) ∧ ( 1 − η ) , ψ η ( t ) := 2 ∫ 0 t ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d )) η + f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 d s , t ∈ [ 0 , 1 ] , (90)
obtaining a bilipschitz map ψ η : [ 0 , 1 ] → [ 0 , ψ η ( 1 ) ] \psi_\eta:[0,1]\to[0,\psi_\eta(1)] ψ η : [ 0 , 1 ] → [ 0 , ψ η ( 1 )] . Then we choose competitors for g λ ∗ ( R − , R + ) g_\lambda^*(R^-,R^+) g λ ∗ ( R − , R + ) as follows:
β ^ η ( t ) : = { R − , t ∈ [ − 1 − ψ η ( 1 ) , 0 ] , β ( ψ η − 1 ( t ) ) , t ∈ [ 0 , ψ η ( 1 ) ] , R + , t ∈ [ ψ η ( 1 ) , 1 + ψ η ( 1 ) ] ; v ^ η : = { 1 , t ∈ [ − 1 − ψ η ( 1 ) , − 1 ] , ( 1 − η ) − t η , t ∈ [ − 1 , 0 ] , v η ( ψ η − 1 ( t ) ) , t ∈ [ 0 , ψ η ( 1 ) ] , 1 + ( t − ψ η ( 1 ) − 1 ) η , t ∈ [ ψ η ( 1 ) , 1 + ψ η ( 1 ) ] . \widehat{\beta}_\eta(t):=
\begin{cases}
R^-, & t\in[-1-\psi_\eta(1),0],\\
\beta(\psi_\eta^{-1}(t)), & t\in[0,\psi_\eta(1)],\\
R^+, & t\in[\psi_\eta(1),1+\psi_\eta(1)];
\end{cases}
\qquad
\widehat{v}_\eta:=
\begin{cases}
1, & t\in[-1-\psi_\eta(1),-1],\\
(1-\eta)-t\eta, & t\in[-1,0],\\
v_\eta(\psi_\eta^{-1}(t)), & t\in[0,\psi_\eta(1)],\\
1+(t-\psi_\eta(1)-1)\eta, & t\in[\psi_\eta(1),1+\psi_\eta(1)].
\end{cases} β η ( t ) := ⎩ ⎨ ⎧ R − , β ( ψ η − 1 ( t )) , R + , t ∈ [ − 1 − ψ η ( 1 ) , 0 ] , t ∈ [ 0 , ψ η ( 1 )] , t ∈ [ ψ η ( 1 ) , 1 + ψ η ( 1 )] ; v η := ⎩ ⎨ ⎧ 1 , ( 1 − η ) − t η , v η ( ψ η − 1 ( t )) , 1 + ( t − ψ η ( 1 ) − 1 ) η , t ∈ [ − 1 − ψ η ( 1 ) , − 1 ] , t ∈ [ − 1 , 0 ] , t ∈ [ 0 , ψ η ( 1 )] , t ∈ [ ψ η ( 1 ) , 1 + ψ η ( 1 )] .
Namely, setting T η : = ψ η ( 1 ) + 1 T_\eta:=\psi_\eta(1)+1 T η := ψ η ( 1 ) + 1 , we have that ( β ^ η , v ^ η ) ∈ U T η [ R − , R + ] (\widehat{\beta}_\eta,\widehat{v}_\eta)\in\mathcal{U}^{T_\eta}[R^-,R^+] ( β η , v η ) ∈ U T η [ R − , R + ] , and hence
g λ ∗ ( R − , R + ) ≤ F λ T η ( β ^ η , v ^ η ) , (91) g_\lambda^*(R^-,R^+)\leq F_\lambda^{T_\eta}(\widehat{\beta}_\eta,\widehat{v}_\eta), \tag*{(91)} g λ ∗ ( R − , R + ) ≤ F λ T η ( β η , v η ) , (91)
due to the fact that the optimization problems defining g λ ∗ ( R − , R + ) g_\lambda^*(R^-,R^+) g λ ∗ ( R − , R + ) are decreasing in T T T . We now prove that
F λ T η ( β ^ η , v ^ η ) < ∫ 0 1 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d ) ) f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t + o η → 0 ( 1 ) . F_\lambda^{T_\eta}(\widehat{\beta}_\eta,\widehat{v}_\eta)<\int_0^1\sqrt{(1-v)^2+4\lambda\operatorname{dist}^2(\beta,\mathrm{O}(d))}\sqrt{f^2(v)|\beta'|^2+|v'|^2}\,\mathrm{d}t+o_{\eta\to0}(1). F λ T η ( β η , v η ) < ∫ 0 1 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t + o η → 0 ( 1 ) .
Using the change of variable t = ψ η ( s ) t=\psi_\eta(s) t = ψ η ( s ) , it is straightforward that
∫ 0 ψ η ( 1 ) ( f 2 ( v ^ η ) ∣ β ^ η ′ ∣ 2 + ∣ v ^ η ′ ∣ 2 ) d t + ∫ 0 ψ η ( 1 ) ( ( 1 − v ^ η ) 2 4 + λ dist 2 ( β ^ η , O ( d ) ) ) d t = ∫ 0 1 ( f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ) 1 ψ η ′ d s + ∫ 0 1 ( ( 1 − v η ) 2 4 + λ dist 2 ( β , O ( d ) ) ) ψ η ′ d s = 1 2 ∫ 0 1 ( f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ) ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d ) ) η + f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 d s + 1 2 ∫ 0 1 η + f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d ) ) d s ≤ c λ η 2 + ∫ 0 1 f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d ) ) d s , \begin{aligned}
&\int_0^{\psi_\eta(1)}\left(f^2(\widehat{v}_\eta)|\widehat{\beta}_\eta'|^2+|\widehat{v}_\eta'|^2\right)\,\mathrm{d}t+\int_0^{\psi_\eta(1)}\left(\frac{(1-\widehat{v}_\eta)^2}{4}+\lambda\operatorname{dist}^2(\widehat{\beta}_\eta,\mathrm{O}(d))\right)\,\mathrm{d}t\\
&=\int_0^1\left(f^2(v_\eta)|\beta'|^2+|v_\eta'|^2\right)\frac{1}{\psi_\eta'}\,\mathrm{d}s+\int_0^1\left(\frac{(1-v_\eta)^2}{4}+\lambda\operatorname{dist}^2(\beta,\mathrm{O}(d))\right)\psi_\eta'\,\mathrm{d}s\\
&=\frac{1}{2}\int_0^1\left(f^2(v_\eta)|\beta'|^2+|v_\eta'|^2\right)\sqrt{\frac{(1-v_\eta)^2+4\lambda\operatorname{dist}^2(\beta,\mathrm{O}(d))}{\eta+f^2(v_\eta)|\beta'|^2+|v_\eta'|^2}}\,\mathrm{d}s\\
&\quad+\frac{1}{2}\int_0^1\sqrt{\eta+f^2(v_\eta)|\beta'|^2+|v_\eta'|^2}\sqrt{(1-v_\eta)^2+4\lambda\operatorname{dist}^2(\beta,\mathrm{O}(d))}\,\mathrm{d}s\\
&\leq c_\lambda\frac{\sqrt{\eta}}{2}+\int_0^1\sqrt{f^2(v_\eta)|\beta'|^2+|v_\eta'|^2}\sqrt{(1-v_\eta)^2+4\lambda\operatorname{dist}^2(\beta,\mathrm{O}(d))}\,\mathrm{d}s,
\end{aligned} ∫ 0 ψ η ( 1 ) ( f 2 ( v η ) ∣ β η ′ ∣ 2 + ∣ v η ′ ∣ 2 ) d t + ∫ 0 ψ η ( 1 ) ( 4 ( 1 − v η ) 2 + λ dist 2 ( β η , O ( d )) ) d t = ∫ 0 1 ( f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ) ψ η ′ 1 d s + ∫ 0 1 ( 4 ( 1 − v η ) 2 + λ dist 2 ( β , O ( d )) ) ψ η ′ d s = 2 1 ∫ 0 1 ( f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ) η + f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d )) d s + 2 1 ∫ 0 1 η + f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d )) d s ≤ c λ 2 η + ∫ 0 1 f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d )) d s ,
where c λ = 1 + 8 λ d c_\lambda=\sqrt{1+8\lambda d} c λ = 1 + 8 λ d (we used here subadditivity of the square root and the bound dist ( β , O ( d ) ) ≤ 2 d \operatorname{dist}(\beta,\mathrm{O}(d))\leq2\sqrt{d} dist ( β , O ( d )) ≤ 2 d ). Denoting E η : = I T η ∖ [ 0 , ψ η ( 1 ) ] E_\eta:=I_{T_\eta}\setminus[0,\psi_\eta(1)] E η := I T η ∖ [ 0 , ψ η ( 1 )] , we have
∫ E η ( f 2 ( v ^ η ) ∣ β ^ η ′ ∣ 2 + ∣ v ^ η ′ ∣ 2 + ( 1 − v ^ η ) 2 4 + λ dist 2 ( β ^ η , O ( d ) ) ) d t = ∫ E η ( ∣ v ^ η ′ ∣ 2 + ( 1 − v ^ η ) 2 4 ) d t = 2 η 2 + η 2 6 , \begin{aligned}
\int_{E_\eta}\left(f^2(\widehat{v}_\eta)|\widehat{\beta}_\eta'|^2+|\widehat{v}_\eta'|^2+\frac{(1-\widehat{v}_\eta)^2}{4}+\lambda\operatorname{dist}^2(\widehat{\beta}_\eta,\mathrm{O}(d))\right)\,\mathrm{d}t \\
&=\int_{E_\eta}\left(|\widehat{v}_\eta'|^2+\frac{(1-\widehat{v}_\eta)^2}{4}\right)\,\mathrm{d}t=2\eta^2+\frac{\eta^2}{6},
\end{aligned} ∫ E η ( f 2 ( v η ) ∣ β η ′ ∣ 2 + ∣ v η ′ ∣ 2 + 4 ( 1 − v η ) 2 + λ dist 2 ( β η , O ( d )) ) d t = ∫ E η ( ∣ v η ′ ∣ 2 + 4 ( 1 − v η ) 2 ) d t = 2 η 2 + 6 η 2 ,
and hence
F λ T η ( β ^ η , v ^ η ) ≤ c λ η + 3 η 2 + ∫ 0 1 f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d ) ) d t . F_\lambda^{T_\eta}(\widehat{\beta}_\eta,\widehat{v}_\eta)\leq c_\lambda\sqrt{\eta}+3\eta^2+\int_0^1\sqrt{f^2(v_\eta)|\beta'|^2+|v_\eta'|^2}\sqrt{(1-v_\eta)^2+4\lambda\operatorname{dist}^2(\beta,\mathrm{O}(d))}\,\mathrm{d}t. F λ T η ( β η , v η ) ≤ c λ η + 3 η 2 + ∫ 0 1 f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d )) d t .
Since v η ′ = 0 v_\eta'=0 v η ′ = 0 almost everywhere in the set { v η ≠ v } \{v_\eta\ne v\} { v η = v } , we have
∫ 0 1 f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d ) ) d t = ∫ { v η = v } f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d ) ) d t + ∫ { v η ≠ v } f ( 1 − η ) ∣ β ′ ∣ η 2 + 4 λ dist 2 ( β , O ( d ) ) d t . \begin{aligned}
\int_0^1 f^2(v_\eta)|\beta'|^2+|v_\eta'|^2\sqrt{(1-v_\eta)^2+4\lambda\operatorname{dist}^2(\beta,O(d))}\,\mathrm{d}t
&=\int_{\{v_\eta=v\}}\sqrt{f^2(v)|\beta'|^2+|v'|^2}\sqrt{(1-v)^2+4\lambda\operatorname{dist}^2(\beta,O(d))}\,\mathrm{d}t\\
&\quad+\int_{\{v_\eta\ne v\}}f(1-\eta)|\beta'|\sqrt{\eta^2+4\lambda\operatorname{dist}^2(\beta,O(d))}\,\mathrm{d}t.
\end{aligned} ∫ 0 1 f 2 ( v η ) ∣ β ′ ∣ 2 + ∣ v η ′ ∣ 2 ( 1 − v η ) 2 + 4 λ dist 2 ( β , O ( d )) d t = ∫ { v η = v } f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) d t + ∫ { v η = v } f ( 1 − η ) ∣ β ′ ∣ η 2 + 4 λ dist 2 ( β , O ( d )) d t .
Finally, let ω ‾ ( η ) \overline{\omega}(\eta) ω ( η ) be the continuity modulus of 1 − s ∣ log ( 1 − s ) ∣ f ( s ) \frac{1-s}{|\log(1-s)|}f(s) ∣ l o g ( 1 − s ) ∣ 1 − s f ( s ) near 1 1 1 , so that
∣ log ( 1 − s ) ∣ 1 − s ( ℓ − ω ‾ ( η ) ) ≤ f ( s ) ≤ ∣ log ( 1 − s ) ∣ 1 − s ( ℓ + ω ‾ ( η ) ) , s ∈ [ 1 − η , 1 ) . (92) \frac{|\log(1-s)|}{1-s}(\ell-\overline{\omega}(\eta))\le f(s)\le\frac{|\log(1-s)|}{1-s}(\ell+\overline{\omega}(\eta)),\qquad s\in[1-\eta,1). \tag*{(92)} 1 − s ∣ log ( 1 − s ) ∣ ( ℓ − ω ( η )) ≤ f ( s ) ≤ 1 − s ∣ log ( 1 − s ) ∣ ( ℓ + ω ( η )) , s ∈ [ 1 − η , 1 ) . (92)
Since for every a ≥ 0 a\ge0 a ≥ 0 the function s ↦ ∣ log ( 1 − s ) ∣ 1 − s ( 1 − s ) 2 + a = ∣ log ( 1 − s ) ∣ 1 + a ( 1 − s ) − 2 s\mapsto\frac{|\log(1-s)|}{1-s}\sqrt{(1-s)^2+a}=|\log(1-s)|\sqrt{1+a(1-s)^{-2}} s ↦ 1 − s ∣ l o g ( 1 − s ) ∣ ( 1 − s ) 2 + a = ∣ log ( 1 − s ) ∣ 1 + a ( 1 − s ) − 2 is increasing in [ 0 , 1 ) [0,1) [ 0 , 1 ) , and v ≥ 1 − η v\ge1-\eta v ≥ 1 − η in the set { v η ≠ v } \{v_\eta\ne v\} { v η = v } , we obtain that
∫ { v η ≠ v } f ( 1 − η ) ∣ β ′ ∣ η 2 + 4 λ dist 2 ( β , O ( d ) ) d t ≤ ℓ + ω ‾ ( η ) ℓ − ω ‾ ( η ) ∫ { v η ≠ v } f ( v ) ∣ β ′ ∣ ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d ) ) d t ≤ ℓ + ω ‾ ( η ) ℓ − ω ‾ ( η ) ∫ { v η ≠ v } f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d ) ) d t . (93) \begin{aligned}
\int_{\{v_\eta\ne v\}}f(1-\eta)|\beta'|\sqrt{\eta^2+4\lambda\operatorname{dist}^2(\beta,O(d))}\,\mathrm{d}t
&\le\frac{\ell+\overline{\omega}(\eta)}{\ell-\overline{\omega}(\eta)}
\int_{\{v_\eta\ne v\}}f(v)|\beta'|\sqrt{(1-v)^2+4\lambda\operatorname{dist}^2(\beta,O(d))}\,\mathrm{d}t\\
&\le\frac{\ell+\overline{\omega}(\eta)}{\ell-\overline{\omega}(\eta)}
\int_{\{v_\eta\ne v\}}\sqrt{f^2(v)|\beta'|^2+|v'|^2}\sqrt{(1-v)^2+4\lambda\operatorname{dist}^2(\beta,O(d))}\,\mathrm{d}t. \tag*{(93)}
\end{aligned} ∫ { v η = v } f ( 1 − η ) ∣ β ′ ∣ η 2 + 4 λ dist 2 ( β , O ( d )) d t ≤ ℓ − ω ( η ) ℓ + ω ( η ) ∫ { v η = v } f ( v ) ∣ β ′ ∣ ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) d t ≤ ℓ − ω ( η ) ℓ + ω ( η ) ∫ { v η = v } f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) d t . (93)
Summarizing, gathering (4.11) and the above computations we obtain that
g λ ∗ ( R − , R + ) ≤ g ‾ λ ( R − , R + ) + c λ η + 4 η 2 + 2 ω ‾ ( η ) ℓ − ω ‾ ( η ) ∫ 0 1 f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d ) ) d t . (94) \begin{aligned}
g_\lambda^*(R^-,R^+)&\le\overline{g}_\lambda(R^-,R^+)\\
&\quad+c_\lambda\sqrt{\eta}+4\eta^2+\frac{2\overline{\omega}(\eta)}{\ell-\overline{\omega}(\eta)}
\int_0^1\sqrt{f^2(v)|\beta'|^2+|v'|^2}\sqrt{(1-v)^2+4\lambda\operatorname{dist}^2(\beta,O(d))}\,\mathrm{d}t. \tag*{(94)}
\end{aligned} g λ ∗ ( R − , R + ) ≤ g λ ( R − , R + ) + c λ η + 4 η 2 + ℓ − ω ( η ) 2 ω ( η ) ∫ 0 1 f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) d t . (94)
This permits us to conclude by the arbitrariness of η > 0 \eta>0 η > 0 . □
At various steps hereafter we shall need to transport trajectories onto O ( d ) O(d) O ( d ) . We present here the basic computation to estimate the length of the resulting curves.
Lemma 4.3. Let I ⊂ R I\subset\mathbb{R} I ⊂ R be an interval, γ ∈ W 1 , 1 ( I ; M d × d ) \gamma\in W^{1,1}(I;\mathbb{M}^{d\times d}) γ ∈ W 1 , 1 ( I ; M d × d ) be such that dist ( γ ( t ) , O ( d ) ) ≤ 1 / 2 \operatorname{dist}(\gamma(t),O(d))\le1/2 dist ( γ ( t ) , O ( d )) ≤ 1/2 almost everywhere. Then there exist C > 0 C>0 C > 0 independent of γ \gamma γ , and γ ^ ∈ W 1 , 1 ( I ; O ( d ) ) \widehat{\gamma}\in W^{1,1}(I;O(d)) γ ∈ W 1 , 1 ( I ; O ( d )) such that ∣ γ ^ ′ ( t ) ∣ ≤ ( 1 + C dist ( γ ( t ) , O ( d ) ) ) ∣ γ ′ ( t ) ∣ |\widehat{\gamma}'(t)|\le(1+C\operatorname{dist}(\gamma(t),O(d)))|\gamma'(t)| ∣ γ ′ ( t ) ∣ ≤ ( 1 + C dist ( γ ( t ) , O ( d ))) ∣ γ ′ ( t ) ∣ for a.e. t ∈ I t\in I t ∈ I , and γ ^ = γ \widehat{\gamma}=\gamma γ = γ on { γ ∈ O ( d ) } \{\gamma\in O(d)\} { γ ∈ O ( d )} .
Proof. We use, as in the argument leading to (3.21), that ψ ( A ) : = A ( A T A ) − 1 / 2 \psi(A):=A(A^TA)^{-1/2} ψ ( A ) := A ( A T A ) − 1/2 is a smooth map from M d × d ∖ { det = 0 } \mathbb{M}^{d\times d}\setminus\{\det=0\} M d × d ∖ { det = 0 } to O ( d ) O(d) O ( d ) , which is in particular Lipschitz on the set { dist ( ⋅ , O ( d ) ) ≤ 1 / 2 } \{\operatorname{dist}(\cdot,O(d))\le1/2\} { dist ( ⋅ , O ( d )) ≤ 1/2 } . At the same time, for R ∈ O ( d ) R\in O(d) R ∈ O ( d ) one has ψ ( R ) = R \psi(R)=R ψ ( R ) = R , and for any ∣ η ∣ ≤ 1 / 2 |\eta|\le1/2 ∣ η ∣ ≤ 1/2 , with η ∈ M d × d \eta\in\mathbb{M}^{d\times d} η ∈ M d × d , a Taylor expansion gives
ψ ( R + η ) = ( R + η ) ( ( R T + η T ) ( R + η ) ) − 1 / 2 = ( R + η ) ( Id + η T R + R T η + η T η ) − 1 / 2 = R + η − 1 2 R η T R − 1 2 η + O ( ∣ η ∣ 2 ) . (95) \begin{aligned}
\psi(R+\eta)&=(R+\eta)\left((R^T+\eta^T)(R+\eta)\right)^{-1/2}\\
&=(R+\eta)(\operatorname{Id}+\eta^TR+R^T\eta+\eta^T\eta)^{-1/2}\\
&=R+\eta-\frac{1}{2}R\eta^TR-\frac{1}{2}\eta+O(|\eta|^2). \tag*{(95)}
\end{aligned} ψ ( R + η ) = ( R + η ) ( ( R T + η T ) ( R + η ) ) − 1/2 = ( R + η ) ( Id + η T R + R T η + η T η ) − 1/2 = R + η − 2 1 R η T R − 2 1 η + O ( ∣ η ∣ 2 ) . (95)
so that ∣ D ψ ( R ) η ∣ ≤ ∣ η ∣ |D\psi(R)\eta|\le|\eta| ∣ D ψ ( R ) η ∣ ≤ ∣ η ∣ , and hence ∣ D ψ ( R ) ∣ op ≤ 1 |D\psi(R)|_{\operatorname{op}}\le1 ∣ D ψ ( R ) ∣ op ≤ 1 , which by regularity of ψ \psi ψ implies ∣ D ψ ( ξ ) ∣ op ≤ 1 + C dist ( ξ , O ( d ) ) |D\psi(\xi)|_{\operatorname{op}}\le1+C\operatorname{dist}(\xi,O(d)) ∣ D ψ ( ξ ) ∣ op ≤ 1 + C dist ( ξ , O ( d )) , whenever dist ( ξ , O ( d ) ) ≤ 1 / 2 \operatorname{dist}(\xi,O(d))\le1/2 dist ( ξ , O ( d )) ≤ 1/2 . Let γ ^ : = ψ ∘ γ \widehat{\gamma}:=\psi\circ\gamma γ := ψ ∘ γ . Then ∣ γ ^ ′ ( t ) ∣ ≤ ∣ D ψ ( γ ( t ) ) ∣ op ∣ γ ′ ( t ) ∣ |\widehat{\gamma}'(t)|\le|D\psi(\gamma(t))|_{\operatorname{op}}|\gamma'(t)| ∣ γ ′ ( t ) ∣ ≤ ∣ D ψ ( γ ( t )) ∣ op ∣ γ ′ ( t ) ∣ and the proof is concluded. □
We are now in a position to prove the main properties of g λ g_\lambda g λ , collected in the following result.
Proposition 4.4. For every λ ∈ [ 0 , + ∞ ] \lambda\in[0,+\infty] λ ∈ [ 0 , + ∞ ] , the function g λ ∗ : O ( d ) × O ( d ) → [ 0 , + ∞ ] g_\lambda^*: \mathrm{O}(d) \times\mathrm{O}(d) \to[0,+\infty] g λ ∗ : O ( d ) × O ( d ) → [ 0 , + ∞ ] satisfies the following properties:
(i) there exists κ ≥ 1 \kappa\ge1 κ ≥ 1 depending only on d d d and f f f and such that
g λ ∗ ( R − , R + ) ≤ 1 ∧ κ ( ∣ R + − R − ∣ ∣ log ∣ R + − R − ∣ ∣ + ∣ R + − R − ∣ ) , (96) g_\lambda^*(R^-,R^+) \le1 \wedge\kappa\left( |R^+ - R^-| \lvert\log|R^+ - R^-| \rvert+ |R^+ - R^-| \right), \tag*{(96)} g λ ∗ ( R − , R + ) ≤ 1 ∧ κ ( ∣ R + − R − ∣ ∣ log ∣ R + − R − ∣ ∣ + ∣ R + − R − ∣ ) , (96) for every R + , R − ∈ O ( d ) R^+,R^- \in\mathrm{O}(d) R + , R − ∈ O ( d ) , λ ∈ [ 0 , ∞ ] \lambda\in[0,\infty] λ ∈ [ 0 , ∞ ] ;
(ii) for every R ∈ O ( d ) R \in\mathrm{O}(d) R ∈ O ( d ) and for every sequence ( R n ) n ⊂ O ( d ) (R_n)_n \subset\mathrm{O}(d) ( R n ) n ⊂ O ( d ) converging to R R R , with R n ≠ R R_n \ne R R n = R , there exists
lim n → + ∞ g λ ∗ ( R n , R ) ∣ R n − R ∣ ∣ log ∣ R n − R ∣ ∣ = ℓ 2 ; \lim_{n \to+\infty} \frac{g_\lambda^*(R_n,R)}{|R_n-R| \lvert\log|R_n-R| \rvert} = \frac{\ell}{2}; n → + ∞ lim ∣ R n − R ∣ ∣ log ∣ R n − R ∣ ∣ g λ ∗ ( R n , R ) = 2 ℓ ; (iii) g λ ∗ ∈ C 0 ( O ( d ) × O ( d ) ) g_\lambda^* \in C^0(\mathrm{O}(d) \times\mathrm{O}(d)) g λ ∗ ∈ C 0 ( O ( d ) × O ( d )) .
All properties are immediately inherited by g λ g_\lambda g λ .
Proof. Proof of (i). It suffices to prove the bound for λ ∈ [ 0 , ∞ ] \lambda\in[0,\infty] λ ∈ [ 0 , ∞ ] , with κ \kappa κ not depending on λ \lambda λ . For every R − , R + ∈ O ( d ) R^-, R^+ \in\mathrm{O}(d) R − , R + ∈ O ( d ) we set
β ˉ ( t ) : = { R − , t ∈ [ 0 , 1 3 ] , R − + ( R + − R − ) ( 3 t − 1 ) , t ∈ [ 1 3 , 2 3 ] , R + , t ∈ [ 2 3 , 1 ] , v ˉ ( t ) : = { 1 − 3 t , t ∈ [ 0 , 1 3 ] , 0 , t ∈ [ 1 3 , 2 3 ] , 3 t − 2 , t ∈ [ 2 3 , 1 ] . \bar{\beta}(t) :=
\begin{cases}
R^-, & t \in\left[0,\frac{1}{3}\right],\\
R^- + (R^+ - R^-)(3t-1), & t \in\left[\frac{1}{3},\frac{2}{3}\right],\\
R^+, & t \in\left[\frac{2}{3},1\right],
\end{cases}
\qquad
\bar{v}(t) :=
\begin{cases}
1-3t, & t \in\left[0,\frac{1}{3}\right],\\
0, & t \in\left[\frac{1}{3},\frac{2}{3}\right],\\
3t-2, & t \in\left[\frac{2}{3},1\right].
\end{cases} β ˉ ( t ) := ⎩ ⎨ ⎧ R − , R − + ( R + − R − ) ( 3 t − 1 ) , R + , t ∈ [ 0 , 3 1 ] , t ∈ [ 3 1 , 3 2 ] , t ∈ [ 3 2 , 1 ] , v ˉ ( t ) := ⎩ ⎨ ⎧ 1 − 3 t , 0 , 3 t − 2 , t ∈ [ 0 , 3 1 ] , t ∈ [ 3 1 , 3 2 ] , t ∈ [ 3 2 , 1 ] .
Applying Proposition 4.2, and by a direct computation, we obtain
g λ ∗ ( R − , R + ) = g ˉ λ ( R − , R + ) ≤ ∫ 0 1 ( 1 − v ˉ ) 2 + 4 λ dist 2 ( β ˉ , O ( d ) ) f 2 ( v ˉ ) ∣ β ˉ ′ ∣ 2 + ∣ v ˉ ′ ∣ 2 d t = ∫ ( 0 , 1 / 3 ) ∪ ( 2 / 3 , 1 ) ( 1 − v ˉ ) ∣ v ˉ ′ ∣ d t = 1. (97) \begin{aligned}
g_\lambda^*(R^-,R^+) = \bar{g}_\lambda(R^-,R^+) \le\int_0^1 \sqrt{(1-\bar{v})^2+4\lambda\operatorname{dist}^2(\bar{\beta},\mathrm{O}(d))}\sqrt{f^2(\bar{v})|\bar{\beta}'|^2+|\bar{v}'|^2}\,\mathrm{d}t \\
&= \int_{(0,1/3)\cup(2/3,1)} (1-\bar{v})|\bar{v}'|\,\mathrm{d}t = 1. \tag*{(97)}
\end{aligned} g λ ∗ ( R − , R + ) = g ˉ λ ( R − , R + ) ≤ ∫ 0 1 ( 1 − v ˉ ) 2 + 4 λ dist 2 ( β ˉ , O ( d )) f 2 ( v ˉ ) ∣ β ˉ ′ ∣ 2 + ∣ v ˉ ′ ∣ 2 d t = ∫ ( 0 , 1/3 ) ∪ ( 2/3 , 1 ) ( 1 − v ˉ ) ∣ v ˉ ′ ∣ d t = 1. (97)
This completes the proof in the case ∣ R + − R − ∣ ≥ 1 |R^+ - R^-| \ge1 ∣ R + − R − ∣ ≥ 1 .
Assume now ∣ R + − R − ∣ < 1 |R^+ - R^-| < 1 ∣ R + − R − ∣ < 1 . We shall complete the proof of (4.17) using again the fact that g λ ∗ = g ˉ λ g_\lambda^*=\bar{g}_\lambda g λ ∗ = g ˉ λ , and choosing properly the competitors in the optimization problem defining g ˉ λ \bar{g}_\lambda g ˉ λ . Using Lemma 4.3 on the affine map joining R − R^- R − with R + R^+ R + , whose image is contained in a neighbourhood of O ( d ) \mathrm{O}(d) O ( d ) of size ∣ R − − R + ∣ / 2 < 1 / 2 |R^- - R^+|/2 < 1/2 ∣ R − − R + ∣/2 < 1/2 , we produce a curve γ ∈ C 1 ( [ 1 3 , 2 3 ] ; O ( d ) ) \gamma\in C^1\left(\left[\frac{1}{3},\frac{2}{3}\right];\mathrm{O}(d)\right) γ ∈ C 1 ( [ 3 1 , 3 2 ] ; O ( d ) ) such that γ ( 1 3 ) = R − \gamma\left(\frac{1}{3}\right)=R^- γ ( 3 1 ) = R − , γ ( 2 3 ) = R + \gamma\left(\frac{2}{3}\right)=R^+ γ ( 3 2 ) = R + , and
∫ 1 3 2 3 ∣ γ ′ ( t ) ∣ d t ≤ ∣ R − − R + ∣ + C ∣ R − − R + ∣ 2 . \int_{\frac{1}{3}}^{\frac{2}{3}} |\gamma'(t)|\,\mathrm{d}t \le|R^- - R^+| + C|R^- - R^+|^2. ∫ 3 1 3 2 ∣ γ ′ ( t ) ∣ d t ≤ ∣ R − − R + ∣ + C ∣ R − − R + ∣ 2 .
For s ∈ ( 0 , 1 ) s \in(0,1) s ∈ ( 0 , 1 ) chosen below, we set
β ( t ) : = { R − , t ∈ [ 0 , 1 3 ] , γ ( t ) , t ∈ [ 1 3 , 2 3 ] , R + , t ∈ [ 2 3 , 1 ] , v ( t ) : = { 1 − 3 t s 1 / 2 , t ∈ [ 0 , 1 3 ] , 1 − s 1 / 2 , t ∈ [ 1 3 , 2 3 ] , 1 + 3 ( t − 1 ) s 1 / 2 , t ∈ [ 2 3 , 1 ] . \beta(t) :=
\begin{cases}
R^-, & t \in\left[0,\frac{1}{3}\right],\\
\gamma(t), & t \in\left[\frac{1}{3},\frac{2}{3}\right],\\
R^+, & t \in\left[\frac{2}{3},1\right],
\end{cases}
\qquad
v(t) :=
\begin{cases}
1-3ts^{1/2}, & t \in\left[0,\frac{1}{3}\right],\\
1-s^{1/2}, & t \in\left[\frac{1}{3},\frac{2}{3}\right],\\
1+3(t-1)s^{1/2}, & t \in\left[\frac{2}{3},1\right].
\end{cases} β ( t ) := ⎩ ⎨ ⎧ R − , γ ( t ) , R + , t ∈ [ 0 , 3 1 ] , t ∈ [ 3 1 , 3 2 ] , t ∈ [ 3 2 , 1 ] , v ( t ) := ⎩ ⎨ ⎧ 1 − 3 t s 1/2 , 1 − s 1/2 , 1 + 3 ( t − 1 ) s 1/2 , t ∈ [ 0 , 3 1 ] , t ∈ [ 3 1 , 3 2 ] , t ∈ [ 3 2 , 1 ] .
An explicit computation, similar to (97) , gives
g ˉ λ ( R − , R + ) ≤ ∫ ( 0 , 1 / 3 ) ∪ ( 2 / 3 , 1 ) ( 1 − v ) ∣ v ′ ∣ d t + ∫ 1 / 3 2 / 3 s 1 / 2 f ( 1 − s 1 / 2 ) ∣ γ ′ ( t ) ∣ d t = s + ( ∣ R − − R + ∣ + C ∣ R − − R + ∣ 2 ) s 1 / 2 f ( 1 − s 1 / 2 ) \bar{g}_{\lambda}(R^{-},R^{+}) \le\int_{(0,1/3)\cup(2/3,1)} (1-v)|v'|\,\mathrm{d}t+\int_{1/3}^{2/3}s^{1/2}f(1-s^{1/2})|\gamma'(t)|\,\mathrm{d}t
= s+\left(|R^{-}-R^{+}|+C|R^{-}-R^{+}|^{2}\right)s^{1/2}f(1-s^{1/2}) g ˉ λ ( R − , R + ) ≤ ∫ ( 0 , 1/3 ) ∪ ( 2/3 , 1 ) ( 1 − v ) ∣ v ′ ∣ d t + ∫ 1/3 2/3 s 1/2 f ( 1 − s 1/2 ) ∣ γ ′ ( t ) ∣ d t = s + ( ∣ R − − R + ∣ + C ∣ R − − R + ∣ 2 ) s 1/2 f ( 1 − s 1/2 )
and hence choosing s = ∣ R − − R + ∣ s=|R^{-}-R^{+}| s = ∣ R − − R + ∣ gives
g ˉ λ ( R − , R + ) ≤ ∣ R − − R + ∣ + ∣ R − − R + ∣ 3 / 2 ( 1 + C ∣ R − − R + ∣ ) f ( 1 − ∣ R − − R + ∣ 1 / 2 ) . (98) \bar{g}_{\lambda}(R^{-},R^{+})\le|R^{-}-R^{+}|+|R^{-}-R^{+}|^{3/2}(1+C|R^{-}-R^{+}|)f(1-|R^{-}-R^{+}|^{1/2}). \tag*{(98)} g ˉ λ ( R − , R + ) ≤ ∣ R − − R + ∣ + ∣ R − − R + ∣ 3/2 ( 1 + C ∣ R − − R + ∣ ) f ( 1 − ∣ R − − R + ∣ 1/2 ) . (98)
The conclusion then follows recalling (1.8) and Proposition 4.2. Proof of (ii). Let ( R n ) n ⊂ O ( d ) (R_n)_n\subset\mathrm{O}(d) ( R n ) n ⊂ O ( d ) converging to R R R be given. Notice that R n R_n R n , R R R belong to the same connected component of O ( d ) \mathrm{O}(d) O ( d ) , say S O ( d ) \mathrm{SO}(d) SO ( d ) , namely R n ∈ S O ( d ) R_n\in\mathrm{SO}(d) R n ∈ SO ( d ) for every n ∈ N n\in\mathbb{N} n ∈ N , R ∈ S O ( d ) R\in\mathrm{SO}(d) R ∈ SO ( d ) . We start by proving the lower bound:
lim inf n → + ∞ g λ ( R n , R ) ∣ R n − R ∣ ∣ log ∣ R n − R ∣ ∣ ≥ ℓ 2 . (99) \liminf_{n\to+\infty}\frac{g_{\lambda}(R_n,R)}{|R_n-R||\log|R_n-R||}\ge\frac{\ell}{2}. \tag*{(99)} n → + ∞ lim inf ∣ R n − R ∣∣ log ∣ R n − R ∣∣ g λ ( R n , R ) ≥ 2 ℓ . (99)
By monotonicity, it suffices to prove it for λ = 0 \lambda=0 λ = 0 . If (99) does not hold, then there are δ > 0 \delta>0 δ > 0 and R n ∈ S O ( d ) R_n\in\mathrm{SO}(d) R n ∈ SO ( d ) , with s n : = ∣ R n − R ∣ → 0 s_n:=|R_n-R|\to0 s n := ∣ R n − R ∣ → 0 , and ( β n , v n ) ∈ U [ R n , R ] (\beta_n,v_n)\in\mathcal{U}[R_n,R] ( β n , v n ) ∈ U [ R n , R ] such that
∫ 0 1 ( 1 − v n ) f 2 ( v n ) ∣ β n ′ ∣ 2 + ∣ v n ′ ∣ 2 d t ≤ ℓ − δ 2 s n ∣ log s n ∣ . (100) \int_0^1(1-v_n)\sqrt{f^2(v_n)|\beta'_n|^2+|v'_n|^2}\,\mathrm{d}t\le\frac{\ell-\delta}{2}s_n|\log s_n|. \tag*{(100)} ∫ 0 1 ( 1 − v n ) f 2 ( v n ) ∣ β n ′ ∣ 2 + ∣ v n ′ ∣ 2 d t ≤ 2 ℓ − δ s n ∣ log s n ∣. (100)
For every n n n , since v n ( 0 ) = v n ( 1 ) = 1 v_n(0)=v_n(1)=1 v n ( 0 ) = v n ( 1 ) = 1 one obtains
( 1 − v n ( t ) ) 2 ≤ ∫ 0 1 ( 1 − v n ) ∣ v n ′ ∣ d t ≤ ℓ 2 s n ∣ log s n ∣ , t ∈ ( 0 , 1 ) . (101) (1-v_n(t))^2\le\int_0^1(1-v_n)|v'_n|\,\mathrm{d}t\le\frac{\ell}{2}s_n|\log s_n|,\qquad t\in(0,1). \tag*{(101)} ( 1 − v n ( t ) ) 2 ≤ ∫ 0 1 ( 1 − v n ) ∣ v n ′ ∣ d t ≤ 2 ℓ s n ∣ log s n ∣ , t ∈ ( 0 , 1 ) . (101)
Moreover, using (100) and ∫ 0 1 ∣ β n ′ ∣ d t ≥ s n \int_0^1|\beta'_n|\,\mathrm{d}t\ge s_n ∫ 0 1 ∣ β n ′ ∣ d t ≥ s n , we get
ℓ − δ 2 s n ∣ log s n ∣ ≥ ∫ 0 1 ( 1 − v n ) f ( v n ) ∣ β n ′ ∣ d t ≥ s n min [ 0 , 1 ] ( 1 − v n ) f ( v n ) . (102) \frac{\ell-\delta}{2}s_n|\log s_n|\ge\int_0^1(1-v_n)f(v_n)|\beta'_n|\,\mathrm{d}t\ge s_n\min_{[0,1]}(1-v_n)f(v_n). \tag*{(102)} 2 ℓ − δ s n ∣ log s n ∣ ≥ ∫ 0 1 ( 1 − v n ) f ( v n ) ∣ β n ′ ∣ d t ≥ s n [ 0 , 1 ] min ( 1 − v n ) f ( v n ) . (102)
Using first (102) and then (101) ,
ℓ − δ 2 ≥ min [ 0 , 1 ] ( 1 − v n ) f ( v n ) ∣ log ( 1 − v n ) ∣ ∣ log ( 1 − v n ) ∣ ∣ log s n ∣ ≥ min [ 0 , 1 ] ( 1 − v n ) f ( v n ) ∣ log ( 1 − v n ) ∣ ∣ log ( ℓ 2 s n ∣ log s n ∣ ) ∣ 2 ∣ log s n ∣ . (103) \frac{\ell-\delta}{2}\ge\min_{[0,1]}\frac{(1-v_n)f(v_n)}{|\log(1-v_n)|}\frac{|\log(1-v_n)|}{|\log s_n|}\ge\min_{[0,1]}\frac{(1-v_n)f(v_n)}{|\log(1-v_n)|}\frac{|\log\left(\frac{\ell}{2}s_n|\log s_n|\right)|}{2|\log s_n|}. \tag*{(103)} 2 ℓ − δ ≥ [ 0 , 1 ] min ∣ log ( 1 − v n ) ∣ ( 1 − v n ) f ( v n ) ∣ log s n ∣ ∣ log ( 1 − v n ) ∣ ≥ [ 0 , 1 ] min ∣ log ( 1 − v n ) ∣ ( 1 − v n ) f ( v n ) 2∣ log s n ∣ ∣ log ( 2 ℓ s n ∣ log s n ∣ ) ∣ . (103)
Now we take the limit n → ∞ n\to\infty n → ∞ . By (101) , v n → 1 v_n\to1 v n → 1 uniformly. Therefore (recalling (1.8)) the first fraction converges to ℓ \ell ℓ . At the same time, since s n → 0 s_n\to0 s n → 0 the second fraction converges to 1 / 2 1/2 1/2 . This gives a contradiction, since we assumed δ > 0 \delta>0 δ > 0 . This proves (99) and then the lower bound.
The upper bound follows immediately using first (98) and then (1.8). Precisely, with s n : = ∣ R n − R ∣ → 0 s_n:=|R_n-R|\to0 s n := ∣ R n − R ∣ → 0 ,
lim sup n → + ∞ g λ ( R n , R ) ∣ R n − R ∣ ∣ log ∣ R n − R ∣ ∣ ≤ lim sup n → + ∞ s n + ( 1 + C s n ) s n 3 / 2 f ( 1 − s n 1 / 2 ) s n ∣ log s n ∣ = lim sup n → + ∞ s n 1 / 2 f ( 1 − s n 1 / 2 ) 2 ∣ log s n 1 / 2 ∣ = ℓ 2 . \limsup_{n\to+\infty}\frac{g_{\lambda}(R_n,R)}{|R_n-R||\log|R_n-R||}\le\limsup_{n\to+\infty}\frac{s_n+(1+Cs_n)s_n^{3/2}f(1-s_n^{1/2})}{s_n|\log s_n|}
=\limsup_{n\to+\infty}\frac{s_n^{1/2}f(1-s_n^{1/2})}{2|\log s_n^{1/2}|}=\frac{\ell}{2}. n → + ∞ lim sup ∣ R n − R ∣∣ log ∣ R n − R ∣∣ g λ ( R n , R ) ≤ n → + ∞ lim sup s n ∣ log s n ∣ s n + ( 1 + C s n ) s n 3/2 f ( 1 − s n 1/2 ) = n → + ∞ lim sup 2∣ log s n 1/2 ∣ s n 1/2 f ( 1 − s n 1/2 ) = 2 ℓ .
Since we used only the bound in (98) , which does not depend on λ \lambda λ , the limit is uniform in λ \lambda λ , and therefore (ii) holds also for λ = ∞ \lambda=\infty λ = ∞ .
Proof of (iii). This follows by combining (96) , which guarantees the continuity at the origin, and the subadditivity stated in Proposition 4.1(v). □
The case λ = ∞ \lambda= \infty λ = ∞ requires a separate treatment for construction of a recovery sequence leading to the upper bound (see § 5.2 ). If R + R^{+} R + and R − R^{-} R − have the same determinant, one can find admissible paths joining them that stay within O ( d ) O(d) O ( d ) . If not, however, such curves do not exist, therefore a different construction is needed. The approximating admissible paths are actually simpler, and can be constructed explicitly, however they depend on δ ε \delta_{\varepsilon} δ ε , therefore in the global construction one cannot take the limit ε → 0 \varepsilon\to0 ε → 0 before the limit T → ∞ T \to\infty T → ∞ . The next Lemma presents this construction.
Lemma 4.5. Assume R + , R − ∈ O ( d ) R^{+}, R^{-} \in O(d) R + , R − ∈ O ( d ) .
(i) If det R + ≠ det R − \det R^{+} \ne\det R^{-} det R + = det R − then there is no ( β , v ) ∈ U T [ R − , R + ] (\beta,v) \in\mathcal{U}^{T}[R^{-},R^{+}] ( β , v ) ∈ U T [ R − , R + ] such that β ( t ) ∈ O ( d ) \beta(t) \in O(d) β ( t ) ∈ O ( d ) almost everywhere,
(ii) If g ∞ ∗ ( R − , R + ) < 1 g_{\infty}^{*}(R^{-},R^{+}) < 1 g ∞ ∗ ( R − , R + ) < 1 then det R + = det R − \det R^{+} = \det R^{-} det R + = det R − and there are sequences ( β T , v T ) ∈ U T [ R − , R + ] (\beta_{T},v_{T}) \in\mathcal{U}^{T}[R^{-},R^{+}] ( β T , v T ) ∈ U T [ R − , R + ] such that β T ( t ) ∈ O ( d ) \beta_{T}(t) \in O(d) β T ( t ) ∈ O ( d ) almost everywhere,
lim T → ∞ F T ( β T , v T ) ≤ g ∞ ∗ ( R − , R + ) , (104) \lim_{T\to\infty} F^{T}(\beta_{T},v_{T}) \le g_{\infty}^{*}(R^{-},R^{+}), \tag*{(104)} T → ∞ lim F T ( β T , v T ) ≤ g ∞ ∗ ( R − , R + ) , (104) and
∥ β T ′ ∥ L 2 ( ( − T , T ) ) 2 ≤ C ( R + , R − ) . (105) \|\beta_{T}'\|_{L^{2}((-T,T))}^{2} \le C(R^{+},R^{-}). \tag*{(105)} ∥ β T ′ ∥ L 2 (( − T , T )) 2 ≤ C ( R + , R − ) . (105) (iii) For any T , λ ≥ 1 T,\lambda\ge1 T , λ ≥ 1 there exists ( β T , λ , v T , λ ) ∈ U T [ R − , R + ] (\beta_{T,\lambda},v_{T,\lambda}) \in\mathcal{U}^{T}[R^{-},R^{+}] ( β T , λ , v T , λ ) ∈ U T [ R − , R + ] such that
F λ T ( β T , λ , v T , λ ) ≤ g ∞ ∗ ( R − , R + ) + ω ( T ) + ω ( λ ) (106) F_{\lambda}^{T}(\beta_{T,\lambda},v_{T,\lambda}) \le g_{\infty}^{*}(R^{-},R^{+})+\omega(T)+\omega(\lambda) \tag*{(106)} F λ T ( β T , λ , v T , λ ) ≤ g ∞ ∗ ( R − , R + ) + ω ( T ) + ω ( λ ) (106) and
∥ β T , λ ′ ∥ L 2 ( ( − T , T ) ) 2 ≤ C λ T , (107) \|\beta_{T,\lambda}'\|_{L^{2}((-T,T))}^{2} \le C\lambda T, \tag*{(107)} ∥ β T , λ ′ ∥ L 2 (( − T , T )) 2 ≤ C λ T , (107) with ω ( s ) → 0 \omega(s) \to0 ω ( s ) → 0 as s → + ∞ s \to+\infty s → + ∞ .
In the upper bound for λ = ∞ \lambda= \infty λ = ∞ we shall use (iii) with sequences λ ε = ε / δ ε \lambda_{\varepsilon} = \varepsilon/\delta_{\varepsilon} λ ε = ε / δ ε and T ε → + ∞ T_{\varepsilon} \to+\infty T ε → + ∞ . One can easily also obtain that (4.27) is an equality, this is however not needed.
Proof. Proof of (i). Since β ∈ H 1 \beta\in H^{1} β ∈ H 1 , it has a continuous representative, and so does det β \det\beta det β . A continuous function which takes values in { ± 1 } \{\pm1\} { ± 1 } is constant.
Proof of (ii). For every λ ≥ 1 \lambda\ge1 λ ≥ 1 we pick T λ T_{\lambda} T λ and then ( β λ , v λ ) ∈ U T λ [ R − , R + ] (\beta_{\lambda},v_{\lambda}) \in\mathcal{U}^{T_{\lambda}}[R^{-},R^{+}] ( β λ , v λ ) ∈ U T λ [ R − , R + ] so that
F λ T λ ( β λ , v λ ) ≤ g λ ∗ ( R − , R + ) + 1 λ . (108) F_{\lambda}^{T_{\lambda}}(\beta_{\lambda},v_{\lambda}) \le g_{\lambda}^{*}(R^{-},R^{+})+\frac{1}{\lambda}. \tag*{(108)} F λ T λ ( β λ , v λ ) ≤ g λ ∗ ( R − , R + ) + λ 1 . (108)
We can assume that β λ \beta_{\lambda} β λ and v λ v_{\lambda} v λ are continuous. We estimate, using that v λ ( ± T λ ) = 1 v_{\lambda}(\pm T_{\lambda})=1 v λ ( ± T λ ) = 1 and the fundamental theorem of calculus on ( − T λ , t ) (-T_{\lambda},t) ( − T λ , t ) and on ( t , T λ ) (t,T_{\lambda}) ( t , T λ ) ,
max I T λ ( 1 − v λ ) 2 ≤ ∫ − T λ T λ ∣ 1 − v λ ∣ ∣ v λ ′ ∣ d t ≤ g λ ∗ ( R − , R + ) + 1 λ . (109) \max_{I_{T_{\lambda}}}(1-v_{\lambda})^{2} \le\int_{-T_{\lambda}}^{T_{\lambda}} |1-v_{\lambda}|\,|v_{\lambda}'|\,dt \le g_{\lambda}^{*}(R^{-},R^{+})+\frac{1}{\lambda}. \tag*{(109)} I T λ max ( 1 − v λ ) 2 ≤ ∫ − T λ T λ ∣1 − v λ ∣ ∣ v λ ′ ∣ d t ≤ g λ ∗ ( R − , R + ) + λ 1 . (109)
The right-hand side converges to g ∞ ∗ ( R − , R + ) < 1 g_{\infty}^{*}(R^{-},R^{+})<1 g ∞ ∗ ( R − , R + ) < 1 for λ → ∞ \lambda\to\infty λ → ∞ . Setting δ : = 1 4 ( 1 − g ∞ ∗ ( R − , R + ) ) > 0 \delta:= \frac{1}{4}(1-g_{\infty}^{*}(R^{-},R^{+}))>0 δ := 4 1 ( 1 − g ∞ ∗ ( R − , R + )) > 0 , for λ \lambda λ large we have
( 1 − v λ ( t ) ) 2 ≤ 1 − 2 δ for all t ∈ I T λ . (110) (1-v_{\lambda}(t))^{2} \le1-2\delta\qquad\text{for all } t \in I_{T_{\lambda}}. \tag*{(110)} ( 1 − v λ ( t ) ) 2 ≤ 1 − 2 δ for all t ∈ I T λ . (110)
By monotonicity of f f f , this implies f ( v λ ) ≥ f ( δ ) f(v_{\lambda}) \ge f(\delta) f ( v λ ) ≥ f ( δ ) everywhere. Therefore (4.29) implies
∫ − T λ T λ ( ∣ β λ ′ ∣ 2 + ∣ v λ ′ ∣ 2 ) d t ≤ C , (111) \int_{-T_{\lambda}}^{T_{\lambda}} \left(|\beta_{\lambda}'|^{2}+|v_{\lambda}'|^{2}\right)\,dt \le C, \tag*{(111)} ∫ − T λ T λ ( ∣ β λ ′ ∣ 2 + ∣ v λ ′ ∣ 2 ) d t ≤ C , (111)
where C > 0 C>0 C > 0 may depend on R − , R + R^{-},R^{+} R − , R + via δ \delta δ , but not on λ \lambda λ . By the same computation as in (109) , using that dist ( β λ ( ± T λ ) , O ( d ) ) = 0 \operatorname{dist}(\beta_{\lambda}(\pm T_{\lambda}),O(d))=0 dist ( β λ ( ± T λ ) , O ( d )) = 0 ,
2 λ 1 / 2 f ( δ ) max t ∈ I T λ dist 2 ( β λ ( t ) , O ( d ) ) ≤ ∫ − T λ T λ 2 λ 1 / 2 f ( v λ ) dist ( β λ , O ( d ) ) ∣ β λ ′ ∣ d t ≤ g λ ∗ ( R − , R + ) + 1 λ + C ≤ C + 1 , (112) \begin{aligned}
2\lambda^{1/2}f(\delta)\max_{t\in I_{T_{\lambda}}}\operatorname{dist}^{2}(\beta_{\lambda}(t),O(d))\leq\int_{-T_{\lambda}}^{T_{\lambda}}2\lambda^{1/2}f(v_{\lambda})\operatorname{dist}(\beta_{\lambda},O(d))|\beta_{\lambda}'|\,dt \\
&\leq g_{\lambda}^{*}(R^{-},R^{+})+\frac{1}{\lambda}+C\leq C+1, \tag*{(112)}
\end{aligned} 2 λ 1/2 f ( δ ) t ∈ I T λ max dist 2 ( β λ ( t ) , O ( d )) ≤ ∫ − T λ T λ 2 λ 1/2 f ( v λ ) dist ( β λ , O ( d )) ∣ β λ ′ ∣ d t ≤ g λ ∗ ( R − , R + ) + λ 1 + C ≤ C + 1 , (112)
therefore ∥ dist ( β λ , O ( d ) ) ∥ L ∞ ≤ C λ − 1 / 4 \|\operatorname{dist}(\beta_{\lambda},O(d))\|_{L^{\infty}}\leq C\lambda^{-1/4} ∥ dist ( β λ , O ( d )) ∥ L ∞ ≤ C λ − 1/4 (recall that C C C depends on δ \delta δ but not on λ \lambda λ ). Let β ^ λ \widehat{\beta}_{\lambda} β λ be the result of Lemma 4.3 applied to β λ \beta_{\lambda} β λ , which obeys
∣ β ^ λ ′ ∣ ≤ ( 1 + C dist ( β λ , O ( d ) ) ) ∣ β λ ′ ∣ . (113) |\widehat{\beta}_{\lambda}'|\leq(1+C\operatorname{dist}(\beta_{\lambda},O(d)))|\beta_{\lambda}'|. \tag*{(113)} ∣ β λ ′ ∣ ≤ ( 1 + C dist ( β λ , O ( d ))) ∣ β λ ′ ∣. (113)
Therefore, by (111) , β ^ λ ∈ H 1 ( I T λ ; O ( d ) ) \widehat{\beta}_{\lambda}\in H^{1}(I_{T_{\lambda}};O(d)) β λ ∈ H 1 ( I T λ ; O ( d )) , and hence, by (i) det R + = det R − \det R^{+}=\det R^{-} det R + = det R − . Moreover, it holds
F λ T λ ( β ^ λ , v λ ) ≤ ( 1 + C λ 1 / 4 ) 2 F λ T λ ( β λ , v λ ) ≤ g λ ∗ ( R − , R + ) + C λ 1 / 4 + 1 λ , (114) F_{\lambda}^{T_{\lambda}}(\widehat{\beta}_{\lambda},v_{\lambda})\leq\left(1+\frac{C}{\lambda^{1/4}}\right)^{2}F_{\lambda}^{T_{\lambda}}(\beta_{\lambda},v_{\lambda})\leq g_{\lambda}^{*}(R^{-},R^{+})+\frac{C}{\lambda^{1/4}}+\frac{1}{\lambda}, \tag*{(114)} F λ T λ ( β λ , v λ ) ≤ ( 1 + λ 1/4 C ) 2 F λ T λ ( β λ , v λ ) ≤ g λ ∗ ( R − , R + ) + λ 1/4 C + λ 1 , (114)
and, taking λ → ∞ \lambda\to\infty λ → ∞ , we obtain
lim sup λ → ∞ F λ T λ ( β ^ λ , v λ ) ≤ g ∞ ∗ ( R − , R + ) . (115) \limsup_{\lambda\to\infty}F_{\lambda}^{T_{\lambda}}(\widehat{\beta}_{\lambda},v_{\lambda})\leq g_{\infty}^{*}(R^{-},R^{+}). \tag*{(115)} λ → ∞ lim sup F λ T λ ( β λ , v λ ) ≤ g ∞ ∗ ( R − , R + ) . (115)
By monotonicity in T T T , the same holds along any sequence T → ∞ T\to\infty T → ∞ .
Proof of (iii): If g ∞ ∗ ( R − , R + ) < 1 g_{\infty}^{*}(R^{-},R^{+})<1 g ∞ ∗ ( R − , R + ) < 1 , it follows from (ii) with ( β T , λ , v T , λ ) = ( β T , v T ) (\beta_{T,\lambda},v_{T,\lambda})=(\beta_{T},v_{T}) ( β T , λ , v T , λ ) = ( β T , v T ) .
If not, we construct explicitly an appropriate sequence ( β T , λ , v T , λ ) ∈ U T [ R − , R + ] (\beta_{T,\lambda},v_{T,\lambda})\in\mathcal{U}^{T}[R^{-},R^{+}] ( β T , λ , v T , λ ) ∈ U T [ R − , R + ] , for any T > 4 T>4 T > 4 . Fix η ∈ ( 0 , 1 ) \eta\in(0,1) η ∈ ( 0 , 1 ) , chosen below depending on λ \lambda λ and T T T , and set
v T , λ ( t ) : = { 0 , if ∣ t ∣ < η , 1 − e − ( ∣ t ∣ − η ) / 2 , if η ≤ ∣ t ∣ < T − 2 + η , 1 − e − ( T − 2 ) / 2 ( T − 1 + η − ∣ t ∣ ) , if T − 2 + η ≤ ∣ t ∣ < T − 1 + η , 1 , if ∣ t ∣ ≥ T − 1 + η . (116) v_{T,\lambda}(t):=
\begin{cases}
0, & \text{if } |t|<\eta,\\
1-e^{-(|t|-\eta)/2}, & \text{if } \eta\leq|t|<T-2+\eta,\\
1-e^{-(T-2)/2}(T-1+\eta-|t|), & \text{if } T-2+\eta\leq|t|<T-1+\eta,\\
1, & \text{if } |t|\geq T-1+\eta.
\end{cases} \tag*{(116)} v T , λ ( t ) := ⎩ ⎨ ⎧ 0 , 1 − e − ( ∣ t ∣ − η ) /2 , 1 − e − ( T − 2 ) /2 ( T − 1 + η − ∣ t ∣ ) , 1 , if ∣ t ∣ < η , if η ≤ ∣ t ∣ < T − 2 + η , if T − 2 + η ≤ ∣ t ∣ < T − 1 + η , if ∣ t ∣ ≥ T − 1 + η . (116)
Correspondingly, we set
β T , λ ( t ) : = { R − , if t ≤ − η , η + t 2 η R + + η − t 2 η R − , if ∣ t ∣ < η , R + , if t ≥ η , (117) \beta_{T,\lambda}(t):=
\begin{cases}
R^{-}, & \text{if } t\leq-\eta,\\
\frac{\eta+t}{2\eta}R^{+}+\frac{\eta-t}{2\eta}R^{-}, & \text{if } |t|<\eta,\\
R^{+}, & \text{if } t\geq\eta,
\end{cases} \tag*{(117)} β T , λ ( t ) := ⎩ ⎨ ⎧ R − , 2 η η + t R + + 2 η η − t R − , R + , if t ≤ − η , if ∣ t ∣ < η , if t ≥ η , (117)
so that β T , λ ′ ( t ) = 0 \beta_{T,\lambda}'(t)=0 β T , λ ′ ( t ) = 0 for ∣ t ∣ ≥ η |t|\geq\eta ∣ t ∣ ≥ η , and ( β T , λ , v T , λ ) ∈ U T [ R − , R + ] (\beta_{T,\lambda},v_{T,\lambda})\in\mathcal{U}^{T}[R^{-},R^{+}] ( β T , λ , v T , λ ) ∈ U T [ R − , R + ] . We estimate
∫ − T T ( 1 − v T , λ ) 2 4 + ∣ v T , λ ′ ∣ 2 d t ≤ 2 η 4 + 2 ∫ 0 ∞ e − t 4 + e − t 4 d t + 2 e − ( T − 2 ) ( 1 4 + 1 ) \int_{-T}^{T}\frac{(1-v_{T,\lambda})^{2}}{4}+|v_{T,\lambda}'|^{2}\,dt\leq\frac{2\eta}{4}+2\int_{0}^{\infty}\frac{e^{-t}}{4}+\frac{e^{-t}}{4}\,dt+2e^{-(T-2)}\left(\frac{1}{4}+1\right) ∫ − T T 4 ( 1 − v T , λ ) 2 + ∣ v T , λ ′ ∣ 2 d t ≤ 4 2 η + 2 ∫ 0 ∞ 4 e − t + 4 e − t d t + 2 e − ( T − 2 ) ( 4 1 + 1 )
≤ 1 + 3 e − T / 2 + η 2 ≤ g ∞ ∗ ( R − , R + ) + 3 e − T / 2 + η 2 , (118) \leq1+3e^{-T/2}+\frac{\eta}{2}\leq g_{\infty}^{*}(R^{-},R^{+})+3e^{-T/2}+\frac{\eta}{2}, \tag*{(118)} ≤ 1 + 3 e − T /2 + 2 η ≤ g ∞ ∗ ( R − , R + ) + 3 e − T /2 + 2 η , (118)
f ( v T , λ ) ∣ β T , λ ′ ∣ = 0 f(v_{T,\lambda})|\beta_{T,\lambda}'|=0 f ( v T , λ ) ∣ β T , λ ′ ∣ = 0 almost everywhere,
∫ − T T λ dist 2 ( β T , λ , O ( d ) ) d t ≤ 2 η λ 4 d , (119) \int_{-T}^{T}\lambda\operatorname{dist}^{2}(\beta_{T,\lambda},O(d))\,dt\leq2\eta\lambda\,4d, \tag*{(119)} ∫ − T T λ dist 2 ( β T , λ , O ( d )) d t ≤ 2 η λ 4 d , (119)
and
∥ β T , λ ′ ∥ L 2 ( ( − T , T ) ) 2 = ∣ R + − R − ∣ 2 ( 2 η ) 2 2 η ≤ d η . (120) \|\beta_{T,\lambda}'\|_{L^{2}((-T,T))}^{2}=\frac{|R^{+}-R^{-}|^{2}}{(2\eta)^{2}}2\eta\leq\frac{d}{\eta}. \tag*{(120)} ∥ β T , λ ′ ∥ L 2 (( − T , T )) 2 = ( 2 η ) 2 ∣ R + − R − ∣ 2 2 η ≤ η d . (120)
At this point we choose η : = 1 / ( λ T ) \eta:=1/(\lambda T) η := 1/ ( λ T ) and conclude the proof. □
Proof of Γ \Gamma Γ -convergence
In this last section we show the Γ \Gamma Γ -convergence result by proving separately the lower bound for F ~ ε \widetilde{\mathcal{F}}_\varepsilon F ε and the upper bound for F ε \mathcal{F}_\varepsilon F ε (recall (51) and (52) ).
Lower bound
This section is devoted to the proof of the following lower bound result (see (LB)′ ' ′ in Theorem 2.15 ).
Theorem 5.1. For any ( u ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d / G × [ 0 , 1 ] ) (u_\varepsilon,v_\varepsilon)_\varepsilon\subset L^1(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}\times[0,1]) ( u ε , v ε ) ε ⊂ L 1 ( Ω ; M d × d / G × [ 0 , 1 ]) and u ∈ S B V 0 ( Ω ; O ( d ) / G ) u\in SBV_0(\Omega;O(d)/\mathcal{G}) u ∈ S B V 0 ( Ω ; O ( d ) / G ) satisfying
u ε → u in L 1 ( Ω ; M d × d / G ) , v ε → 1 in L 1 ( Ω ) . u_\varepsilon\to u \text{ in } L^1(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}), \qquad v_\varepsilon\to1 \text{ in } L^1(\Omega). u ε → u in L 1 ( Ω ; M d × d / G ) , v ε → 1 in L 1 ( Ω ) . it holds
F ~ λ ( u , 1 ) ≤ lim inf ε → 0 F ~ ε ( u ε , v ε ) . \widetilde{\mathcal{F}}^\lambda(u,1)\leq\liminf_{\varepsilon\to0}\widetilde{\mathcal{F}}_\varepsilon(u_\varepsilon,v_\varepsilon). F λ ( u , 1 ) ≤ ε → 0 lim inf F ε ( u ε , v ε ) .
The proof uses first a blow up argument to reduce to the case of a flat interface with the limiting u u u taking only two values and then a slicing argument for functions in H 1 ( Ω ; M d × d / G ) H^1(\Omega;\mathbb{M}^{d\times d}/\mathcal{G}) H 1 ( Ω ; M d × d / G ) to reduce to a one dimensional problem in the interval I : = ( − 1 2 , 1 2 ) I:=\left(-\frac{1}{2},\frac{1}{2}\right) I := ( − 2 1 , 2 1 ) . Being in dimension one allows us to exploit the relation between H 1 ( I ; M d × d / G ) H^1(I;\mathbb{M}^{d\times d}/\mathcal{G}) H 1 ( I ; M d × d / G ) and H 1 ( I ; M d × d ) H^1(I;\mathbb{M}^{d\times d}) H 1 ( I ; M d × d ) discussed in Lemma 2.5. The main additional difficulty in the treatment of these one dimensional sections, which is handled in a sequence of lemmas before the main proof, is the fact that the functionals F ~ ε \widetilde{\mathcal{F}}_\varepsilon F ε are not invariant under scaling, because the truncation with M ε M_\varepsilon M ε of the prefactor f ε f_\varepsilon f ε scales differently from the rest. Indeed the surface energy g λ g_\lambda g λ defined in Section 4 does not contain any truncation.
To deal with this, we introduce an intermediate step with a further truncation in the prefactor and for every j ∈ N j\in\mathbb{N} j ∈ N we replace f f f by
h j ( s ) : = j s 1 − s ∧ f ( s ) , s ∈ [ 0 , 1 ) , (121) h_j(s):=\frac{js}{1-s}\wedge f(s), \qquad s\in[0,1), \tag*{(121)} h j ( s ) := 1 − s j s ∧ f ( s ) , s ∈ [ 0 , 1 ) , (121)
and we introduce the corresponding surface energy densities
g j , λ ( R − , R + ) : = inf { ∫ − 1 2 1 2 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d ) ) h j 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t : ( β , v ) ∈ U 1 / 2 [ R − , R + ] } (122) g_{j,\lambda}(R^-,R^+):=\inf\left\{\int_{-\frac{1}{2}}^{\frac{1}{2}}\sqrt{(1-v)^2+4\lambda\operatorname{dist}^2(\beta,O(d))}\sqrt{h_j^2(v)|\beta'|^2+|v'|^2}\,dt:
(\beta,v)\in U^{1/2}[R^-,R^+]\right\} \tag*{(122)} g j , λ ( R − , R + ) := inf { ∫ − 2 1 2 1 ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) h j 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t : ( β , v ) ∈ U 1/2 [ R − , R + ] } (122)
for λ ∈ [ 0 , + ∞ ) \lambda\in[0,+\infty) λ ∈ [ 0 , + ∞ ) , and
g j , ∞ ( R − , R + ) : = sup λ > 0 g j , λ ( R − , R + ) . (123) g_{j,\infty}(R^-,R^+):=\sup_{\lambda>0}g_{j,\lambda}(R^-,R^+). \tag*{(123)} g j , ∞ ( R − , R + ) := λ > 0 sup g j , λ ( R − , R + ) . (123)
The set of admissible pairs U 1 / 2 [ R − , R + ] U^{1/2}[R^-,R^+] U 1/2 [ R − , R + ] is defined in (81) . These are good approximations of g λ g_\lambda g λ as shown in Lemma 5.2 below. The reason for doing this is that it permits to replace the phase field v ε v_\varepsilon v ε with a truncated version v ~ ε \widetilde v_\varepsilon v ε with a small error and, at the same time, to remove the truncation from the prefactor.
Lemma 5.2. For any λ ∈ [ 0 , + ∞ ] \lambda\in[0,+\infty] λ ∈ [ 0 , + ∞ ] , R + , R − ∈ O ( d ) R^+,R^-\in O(d) R + , R − ∈ O ( d ) , the following holds:
g λ ( R − , R + ) ≤ lim j → + ∞ g j , λ ( R − , R + ) . (124) g_\lambda(R^-,R^+)\leq\lim_{j\to+\infty}g_{j,\lambda}(R^-,R^+). \tag*{(124)} g λ ( R − , R + ) ≤ j → + ∞ lim g j , λ ( R − , R + ) . (124)
Proof. The proof follows closely that of [19 Proposition 7.3 (iii)] . We sketch here the main steps in the case λ ∈ [ 0 , + ∞ ) \lambda\in[0,+\infty) λ ∈ [ 0 , + ∞ ) , being the case λ = + ∞ \lambda=+\infty λ = + ∞ obtained by taking the supremum over λ > 0 \lambda>0 λ > 0 . For fixed R − , R + ∈ O ( d ) R^-,R^+\in O(d) R − , R + ∈ O ( d ) , for any Borel set E ⊂ R E\subset\mathbb{R} E ⊂ R , we consider the functionals defined in H 1 ( J ; M d × d ) × H 1 ( J ; [ 0 , 1 ] ) H^1(J;\mathbb{M}^{d\times d})\times H^1(J;[0,1]) H 1 ( J ; M d × d ) × H 1 ( J ; [ 0 , 1 ]) for any interval J J J that contains E E E by
H j , λ ( β , v ; E ) : = ∫ E ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d ) ) h j 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t , (125) H_{j,\lambda}(\beta,v;E):=\int_E\sqrt{(1-v)^2+4\lambda\operatorname{dist}^2(\beta,O(d))}\sqrt{h_j^2(v)|\beta'|^2+|v'|^2}\,dt, \tag*{(125)} H j , λ ( β , v ; E ) := ∫ E ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) h j 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t , (125)
H λ ( β , v ; E ) : = ∫ E ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d ) ) f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t , H_\lambda(\beta,v;E):=\int_E\sqrt{(1-v)^2+4\lambda\operatorname{dist}^2(\beta,O(d))}\sqrt{f^2(v)|\beta'|^2+|v'|^2}\,dt, H λ ( β , v ; E ) := ∫ E ( 1 − v ) 2 + 4 λ dist 2 ( β , O ( d )) f 2 ( v ) ∣ β ′ ∣ 2 + ∣ v ′ ∣ 2 d t ,
so that
g j , λ ( R − , R + ) = inf ( β , v ) ∈ U 1 / 2 [ R − , R + ] H j , λ ( β , v , I ) , g λ ∗ ( R − , R + ) = inf ( β , v ) ∈ U 1 / 2 [ R − , R + ] H λ ( β , v , I ) . g_{j,\lambda}(R^-,R^+) = \inf_{(\beta,v)\in U^{1/2}[R^-,R^+]} H_{j,\lambda}(\beta,v,I), \qquad g^*_\lambda(R^-,R^+) = \inf_{(\beta,v)\in U^{1/2}[R^-,R^+]} H_\lambda(\beta,v,I). g j , λ ( R − , R + ) = ( β , v ) ∈ U 1/2 [ R − , R + ] inf H j , λ ( β , v , I ) , g λ ∗ ( R − , R + ) = ( β , v ) ∈ U 1/2 [ R − , R + ] inf H λ ( β , v , I ) .
First we note that by (5.1) and Prop. 4.4(i),
g j , λ ( R − , R + ) ≤ g j + 1 , λ ( R − , R + ) ≤ g λ ∗ ( R − , R + ) ≤ 1 , g_{j,\lambda}(R^-,R^+) \le g_{j+1,\lambda}(R^-,R^+) \le g^*_\lambda(R^-,R^+) \le1, g j , λ ( R − , R + ) ≤ g j + 1 , λ ( R − , R + ) ≤ g λ ∗ ( R − , R + ) ≤ 1 ,
and hence we have to show that
g λ ( R − , R + ) ≤ sup j g j , λ ( R − , R + ) . (126) g_\lambda(R^-,R^+) \le\sup_j g_{j,\lambda}(R^-,R^+). \tag*{(126)} g λ ( R − , R + ) ≤ j sup g j , λ ( R − , R + ) . (126)
For any j ∈ N j \in\mathbb{N} j ∈ N , let ( β j , v j ) ∈ U 1 / 2 [ R − , R + ] (\beta_j,v_j) \in U^{1/2}[R^-,R^+] ( β j , v j ) ∈ U 1/2 [ R − , R + ] be such that
H j , λ ( β j , v j ; I ) ≤ g j , λ ( R − , R + ) + 1 j . (127) H_{j,\lambda}(\beta_j,v_j;I) \le g_{j,\lambda}(R^-,R^+) + \frac{1}{j}. \tag*{(127)} H j , λ ( β j , v j ; I ) ≤ g j , λ ( R − , R + ) + j 1 . (127)
If lim inf j min [ 0 , 1 ] v j = 0 \liminf_j \min_{[0,1]} v_j=0 lim inf j min [ 0 , 1 ] v j = 0 , then the same argument as in (109) easily gives
lim sup j → + ∞ H j , λ ( β j , v j ; I ) ≥ 1 \limsup_{j\to+\infty} H_{j,\lambda}(\beta_j,v_j;I) \ge1 j → + ∞ lim sup H j , λ ( β j , v j ; I ) ≥ 1
which concludes the proof.
Assume now that lim inf j min [ 0 , 1 ] v j > 0 \liminf_j \min_{[0,1]} v_j>0 lim inf j min [ 0 , 1 ] v j > 0 and pick 0 < η < lim inf j min [ 0 , 1 ] v j 0<\eta<\liminf_j \min_{[0,1]}v_j 0 < η < lim inf j min [ 0 , 1 ] v j . We define v j η : = v j ∧ ( 1 − η ) v_j^\eta:=v_j\wedge(1-\eta) v j η := v j ∧ ( 1 − η ) . Extending β j \beta_j β j constant outside I I I and interpolating v j η v_j^\eta v j η linearly in order to achieve the boundary condition 1 1 1 in the interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) , and then using the invariance under reparametrization of H λ H_\lambda H λ , we immediately obtain that
g λ ( R + , R − ) ≤ g λ ∗ ( R + , R − ) ≤ H λ ( β j , v j η ; ( − 1 , 1 ) ) ≤ H λ ( β j , v j η ; I ) + η 2 . (128) g_\lambda(R^+,R^-) \le g^*_\lambda(R^+,R^-) \le H_\lambda(\beta_j,v_j^\eta;(-1,1)) \le H_\lambda(\beta_j,v_j^\eta;I)+\eta^2. \tag*{(128)} g λ ( R + , R − ) ≤ g λ ∗ ( R + , R − ) ≤ H λ ( β j , v j η ; ( − 1 , 1 )) ≤ H λ ( β j , v j η ; I ) + η 2 . (128)
We subdivide I I I in three subsets:
A j : = { t ∈ I : v j ( t ) ≤ 1 − η } , A_j:=\{t\in I:v_j(t)\le1-\eta\}, A j := { t ∈ I : v j ( t ) ≤ 1 − η } ,
B j : = { t ∈ I : 1 − η < v j ( t ) ≤ 1 , h j ( v j ( t ) ) = f ( v j ( t ) ) } , B_j:=\{t\in I:1-\eta<v_j(t)\le1,\ h_j(v_j(t))=f(v_j(t))\}, B j := { t ∈ I : 1 − η < v j ( t ) ≤ 1 , h j ( v j ( t )) = f ( v j ( t ))} ,
C j : = { t ∈ I : 1 − η < v j ( t ) ≤ 1 , h j ( v j ( t ) ) < f ( v j ( t ) ) } . (129) C_j:=\{t\in I:1-\eta<v_j(t)\le1,\ h_j(v_j(t))<f(v_j(t))\}. \tag*{(129)} C j := { t ∈ I : 1 − η < v j ( t ) ≤ 1 , h j ( v j ( t )) < f ( v j ( t ))} . (129)
We start with A j A_j A j . Here v j = v j η v_j=v_j^\eta v j = v j η and since for j j j sufficiently large h j = f h_j=f h j = f in [ η , 1 − η ] [\eta,1-\eta] [ η , 1 − η ] ,
H λ ( β j , v j η ; A j ) = H j , λ ( β j , v j ; A j ) H_\lambda(\beta_j,v_j^\eta;A_j)=H_{j,\lambda}(\beta_j,v_j;A_j) H λ ( β j , v j η ; A j ) = H j , λ ( β j , v j ; A j )
for all j j j large enough. {#eq:5.10}
In order to treat the second set we use the same argument as in (92) and (93) and obtain
H λ ( β j , v j η ; B j ) ≤ ℓ + ω ˉ ( η ) ℓ − ω ˉ ( η ) H j , λ ( β j , v j ; B j ) H_\lambda(\beta_j,v_j^\eta;B_j)\le\frac{\ell+\bar\omega(\eta)}{\ell-\bar\omega(\eta)}H_{j,\lambda}(\beta_j,v_j;B_j) H λ ( β j , v j η ; B j ) ≤ ℓ − ω ˉ ( η ) ℓ + ω ˉ ( η ) H j , λ ( β j , v j ; B j )
for all j j j , ω ˉ ( η ) → 0 \bar\omega(\eta)\to0 ω ˉ ( η ) → 0 as η → 0 \eta\to0 η → 0 . {#eq:5.11}
Finally we observe that if t ∈ C j t\in C_j t ∈ C j then ( 1 − v j ( t ) ) h j ( v j ( t ) ) = j v j ( t ) ≥ j ( 1 − η ) (1-v_j(t))h_j(v_j(t))=jv_j(t)\ge j(1-\eta) ( 1 − v j ( t )) h j ( v j ( t )) = j v j ( t ) ≥ j ( 1 − η ) and therefore
( 1 − v j ( t ) ) h j ( v j ( t ) ) ≥ η f ( 1 − η ) = ( 1 − v j η ) f ( v j η ) (1-v_j(t))h_j(v_j(t))\ge\eta f(1-\eta)=(1-v_j^\eta)f(v_j^\eta) ( 1 − v j ( t )) h j ( v j ( t )) ≥ η f ( 1 − η ) = ( 1 − v j η ) f ( v j η )
for all j ≥ η f ( 1 − η ) 1 − η j\ge\frac{\eta f(1-\eta)}{1-\eta} j ≥ 1 − η η f ( 1 − η ) . Using the same argument as for B j B_j B j we conclude that
H λ ( β j , v j η ; C j ) ≤ H j , λ ( β j , v j ; C j ) H_\lambda(\beta_j,v_j^\eta;C_j)\le H_{j,\lambda}(\beta_j,v_j;C_j) H λ ( β j , v j η ; C j ) ≤ H j , λ ( β j , v j ; C j )
for all j ≥ η f ( 1 − η ) 1 − η j\ge\frac{\eta f(1-\eta)}{1-\eta} j ≥ 1 − η η f ( 1 − η ) . {#eq:5.12}
Then (5.6) follows by (5.8), (5.10), (5.11) (5.12), and (5.7), taking first the limit j → + ∞ j\to+\infty j → + ∞ and then η → 0 \eta\to0 η → 0 . □
Next, we show that in the definition of the energies g j , λ g_{j,\lambda} g j , λ one can also consider test functions such that the boundary conditions are fulfilled only asymptotically.
Lemma 5.3. Fix j ∈ N j \in\mathbb{N} j ∈ N , λ ∈ [ 0 , + ∞ ) \lambda\in[0,+\infty) λ ∈ [ 0 , + ∞ ) , R + , R − ∈ O ( d ) R^{+},R^{-} \in O(d) R + , R − ∈ O ( d ) . Consider a sequence ( β k , v k ) ∈ H 1 ( I ; M d × d × [ 0 , 1 ] ) (\beta_k,v_k) \in H^{1}(I;\mathbb{M}^{d\times d}\times[0,1]) ( β k , v k ) ∈ H 1 ( I ; M d × d × [ 0 , 1 ]) such that
β k → L l o c 1 R − χ ( − 1 2 , 0 ) + R + χ ( 0 , 1 2 ) and v k → 1 in L 1 . (130) \beta_k \xrightarrow{L^{1}_{\mathrm{loc}}} R^{-}\chi_{(-\frac{1}{2},0)}+R^{+}\chi_{(0,\frac{1}{2})}\quad\text{and}\quad v_k\to1\text{ in }L^{1}. \tag*{(130)} β k L loc 1 R − χ ( − 2 1 , 0 ) + R + χ ( 0 , 2 1 ) and v k → 1 in L 1 . (130) Then the functional H j , λ H_{j,\lambda} H j , λ defined in (5.5) obeys
min G + , G − ∈ G g j , λ ( G − R − , G + R + ) ≤ lim inf k → ∞ H j , λ ( β k , v k ; I ) . (131) \min_{G^{+},G^{-}\in\mathcal{G}}g_{j,\lambda}(G^{-}R^{-},G^{+}R^{+})\leq\liminf_{k\to\infty}H_{j,\lambda}(\beta_k,v_k;I). \tag*{(131)} G + , G − ∈ G min g j , λ ( G − R − , G + R + ) ≤ k → ∞ lim inf H j , λ ( β k , v k ; I ) . (131)
Proof. As usual, in order to interpolate the field β k \beta_k β k we need to move v k v_k v k away from 1. Possibly passing to a subsequence, we have pointwise convergence almost everywhere, therefore we can choose z ∈ ( 1 4 , 1 2 ) z\in(\frac{1}{4},\frac{1}{2}) z ∈ ( 4 1 , 2 1 ) and G k ± ∈ G G_k^{\pm}\in\mathcal{G} G k ± ∈ G such that
α k : = ∣ G k + R + − β k ( z ) ∣ + ∣ G k − R − − β k ( − z ) ∣ + ∣ 1 − v k ( z ) ∣ + ∣ 1 − v k ( − z ) ∣ → 0 ; (132) \alpha_k:=\lvert G_k^{+}R^{+}-\beta_k(z)\rvert+\lvert G_k^{-}R^{-}-\beta_k(-z)\rvert+\lvert1-v_k(z)\rvert+\lvert1-v_k(-z)\rvert\to0; \tag*{(132)} α k := ∣ G k + R + − β k ( z )∣ + ∣ G k − R − − β k ( − z )∣ + ∣ 1 − v k ( z )∣ + ∣ 1 − v k ( − z )∣ → 0 ; (132)
for k k k sufficiently large we have h j ( s ) = j s / ( 1 − s ) h_j(s)=js/(1-s) h j ( s ) = j s / ( 1 − s ) for all s ∈ [ 1 − α k , 1 ) s\in[1-\alpha_k,1) s ∈ [ 1 − α k , 1 ) . Therefore on the set { v k ≥ 1 − α k } \{v_k\geq1-\alpha_k\} { v k ≥ 1 − α k } , the integrand in the functional H j , λ ( v k , β k ; I ) H_{j,\lambda}(v_k,\beta_k;I) H j , λ ( v k , β k ; I ) can be rewritten as
1 + 4 λ ( 1 − v k ) 2 dist 2 ( β k , O ( d ) ) j 2 v k 2 ∣ β k ′ ∣ 2 + ( 1 − v k ) 2 ∣ v k ′ ∣ 2 (133) \sqrt{1+\frac{4\lambda}{(1-v_k)^2}\operatorname{dist}^{2}(\beta_k,O(d))}\sqrt{\frac{j^2}{v_k^2}\lvert\beta'_k\rvert^2+(1-v_k)^2\lvert v'_k\rvert^2} \tag*{(133)} 1 + ( 1 − v k ) 2 4 λ dist 2 ( β k , O ( d )) v k 2 j 2 ∣ β k ′ ∣ 2 + ( 1 − v k ) 2 ∣ v k ′ ∣ 2 (133)
and, setting v ^ k : = v k ∧ ( 1 − α k ) \widehat{v}_k:=v_k\wedge(1-\alpha_k) v k := v k ∧ ( 1 − α k ) , one obtains
H j , λ ( β k , v ^ k ; ( − z , z ) ) ≤ H j , λ ( β k , v k ; ( − z , z ) ) . (134) H_{j,\lambda}(\beta_k,\widehat{v}_k;(-z,z))\leq H_{j,\lambda}(\beta_k,v_k;(-z,z)). \tag*{(134)} H j , λ ( β k , v k ; ( − z , z )) ≤ H j , λ ( β k , v k ; ( − z , z )) . (134)
We extend v ^ k \widehat{v}_k v k to equal 1 − α k 1-\alpha_k 1 − α k on the rest of [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , to be one outside ( − 2 , 2 ) (-2,2) ( − 2 , 2 ) , and to be the affine interpolation in-between. We then extend β k \beta_k β k so that β k = G k + R + \beta_k=G_k^{+}R^{+} β k = G k + R + on ( 1 , ∞ ) (1,\infty) ( 1 , ∞ ) , β k = G k − R − \beta_k=G_k^{-}R^{-} β k = G k − R − on ( − ∞ , − 1 ) (-\infty,-1) ( − ∞ , − 1 ) , and the affine interpolation in ( − 1 , − z ) (-1,-z) ( − 1 , − z ) and ( z , 1 ) (z,1) ( z , 1 ) . We observe that both functions are in H 1 H^1 H 1 on ( − 2 , 2 ) (-2,2) ( − 2 , 2 ) . We compute
H j , λ ( β k , v ^ k ; ( z , 1 ) ) ≤ ( 1 − z ) α k 2 + 4 λ α k 2 j 2 α k 2 4 α k 2 = 2 j α k 1 + 4 λ , (135) H_{j,\lambda}(\beta_k,\widehat{v}_k;(z,1))\leq(1-z)\sqrt{\alpha_k^2+4\lambda\alpha_k^2}\sqrt{\frac{j^2}{\alpha_k^2}4\alpha_k^2}=2j\alpha_k\sqrt{1+4\lambda}, \tag*{(135)} H j , λ ( β k , v k ; ( z , 1 )) ≤ ( 1 − z ) α k 2 + 4 λ α k 2 α k 2 j 2 4 α k 2 = 2 j α k 1 + 4 λ , (135)
H j , λ ( β k , v ^ k ; ( 1 , 2 ) ) = ∫ 1 2 α k ( 2 − t ) α k d t = 1 2 α k 2 , (136) H_{j,\lambda}(\beta_k,\widehat{v}_k;(1,2))=\int_1^2\alpha_k(2-t)\alpha_k\,dt=\frac{1}{2}\alpha_k^2, \tag*{(136)} H j , λ ( β k , v k ; ( 1 , 2 )) = ∫ 1 2 α k ( 2 − t ) α k d t = 2 1 α k 2 , (136)
and the same on the other side. Since H j , λ H_{j,\lambda} H j , λ is invariant under reparametrizations we obtain
g j , λ ( G k − R − , G k + R + ) ≤ H j , λ ( β k , v ^ k ; ( − 2 , 2 ) ) ≤ H j , λ ( v k , β k ; I ) + α k 2 + 4 j α k 1 + 4 λ (137) g_{j,\lambda}(G_k^{-}R^{-},G_k^{+}R^{+})\leq H_{j,\lambda}(\beta_k,\widehat{v}_k;(-2,2))\leq H_{j,\lambda}(v_k,\beta_k;I)+\alpha_k^2+4j\alpha_k\sqrt{1+4\lambda} \tag*{(137)} g j , λ ( G k − R − , G k + R + ) ≤ H j , λ ( β k , v k ; ( − 2 , 2 )) ≤ H j , λ ( v k , β k ; I ) + α k 2 + 4 j α k 1 + 4 λ (137)
for all k k k sufficiently large; taking the limit k → ∞ k\to\infty k → ∞ the proof is concluded. □
Lemma 5.4. Let ε k / δ k → λ ∈ [ 0 , + ∞ ] \varepsilon_k/\delta_k\to\lambda\in[0,+\infty] ε k / δ k → λ ∈ [ 0 , + ∞ ] , M k → + ∞ M_k\to+\infty M k → + ∞ , and assume that ( β k , v k ) ∈ H 1 ( I ; M d × d × [ 0 , 1 ] ) (\beta_k,v_k)\in H^1(I;\mathbb{M}^{d\times d}\times[0,1]) ( β k , v k ) ∈ H 1 ( I ; M d × d × [ 0 , 1 ]) , I = [ − 1 / 2 , 1 / 2 ] I=[-1/2,1/2] I = [ − 1/2 , 1/2 ] , obeys
β k → R − χ [ − 1 / 2 , 0 ) + R + χ [ 0 , 1 / 2 ] in L G 1 ( Ω , M d × d ) , v k → 1 in L 1 ( I ) \beta_k\to R^{-}\chi_{[-1/2,0)}+R^{+}\chi_{[0,1/2]}\text{ in }L^1_{\mathcal{G}}(\Omega,\mathbb{M}^{d\times d}),\qquad v_k\to1\text{ in }L^1(I) β k → R − χ [ − 1/2 , 0 ) + R + χ [ 0 , 1/2 ] in L G 1 ( Ω , M d × d ) , v k → 1 in L 1 ( I ) for some R − , R + ∈ O ( d ) R^{-},R^{+}\in O(d) R − , R + ∈ O ( d ) . Then
min G + , G − ∈ G g j , λ ( G − R − , G + R + ) ≤ lim inf k → + ∞ ( ∫ I ( M k 2 ∧ ε k f 2 ( v k ) ) ∣ β k ′ ∣ 2 d t + PF ε k ( v k ; I ) + 1 δ k ∫ I dist 2 ( β k , O ( d ) ) d t ) . (138) \min_{G^{+},G^{-}\in\mathcal{G}}g_{j,\lambda}(G^{-}R^{-},G^{+}R^{+})
\leq\liminf_{k\to+\infty}\left(\int_I\left(M_k^2\wedge\varepsilon_k f^2(v_k)\right)\lvert\beta'_k\rvert^2\,dt+\operatorname{PF}_{\varepsilon_k}(v_k;I)+\frac{1}{\delta_k}\int_I\operatorname{dist}^2(\beta_k,O(d))\,dt\right). \tag*{(138)} G + , G − ∈ G min g j , λ ( G − R − , G + R + ) ≤ k → + ∞ lim inf ( ∫ I ( M k 2 ∧ ε k f 2 ( v k ) ) ∣ β k ′ ∣ 2 d t + PF ε k ( v k ; I ) + δ k 1 ∫ I dist 2 ( β k , O ( d )) d t ) . (138)
Proof. Let us first assume λ < + ∞ \lambda< +\infty λ < + ∞ . For h j h_j h j defined in (5.1), we have
∫ I ( M k 2 ∧ ε k f 2 ( v k ) ) ∣ β k ′ ∣ 2 d t ≥ ∫ I ( M k 2 ∧ ε k h j 2 ( v k ) ) ∣ β k ′ ∣ 2 d t . (139) \int_I \left(M_k^2 \wedge\varepsilon_k f^2(v_k)\right) |\beta_k'|^2 \,\mathrm{d}t \ge\int_I \left(M_k^2 \wedge\varepsilon_k h_j^2(v_k)\right) |\beta_k'|^2 \,\mathrm{d}t. \tag*{(139)} ∫ I ( M k 2 ∧ ε k f 2 ( v k ) ) ∣ β k ′ ∣ 2 d t ≥ ∫ I ( M k 2 ∧ ε k h j 2 ( v k ) ) ∣ β k ′ ∣ 2 d t . (139)
Choose γ k : = 1 − j ε k / M k \gamma_k := 1 - j\sqrt{\varepsilon_k}/M_k γ k := 1 − j ε k / M k and define v ~ k : = v k ∧ ( 1 − γ k ) \widetilde{v}_k := v_k \wedge(1-\gamma_k) v k := v k ∧ ( 1 − γ k ) . Then ε k 1 / 2 h j ( γ k ) ≤ ε k 1 / 2 j γ k / ( 1 − γ k ) ≤ M k \varepsilon_k^{1/2} h_j(\gamma_k) \le\varepsilon_k^{1/2} j\gamma_k/(1-\gamma_k) \le M_k ε k 1/2 h j ( γ k ) ≤ ε k 1/2 j γ k / ( 1 − γ k ) ≤ M k and
∫ I ( M k 2 ∧ ε k h j 2 ( v k ) ) ∣ β k ′ ∣ 2 + ( 1 − v k ) 2 4 ε k + ε k ∣ v k ′ ∣ 2 d t ≥ ∫ I ε k h j 2 ( v ~ k ) ∣ β k ′ ∣ 2 + ( 1 − v ~ k ) 2 4 ε k + ε k ∣ v ~ k ′ ∣ 2 d t − ∫ { v k > γ k } ( 1 − γ k ) 2 4 ε k d t ≥ ∫ I ε k h j 2 ( v ~ k ) ∣ β k ′ ∣ 2 + ( 1 − v ~ k ) 2 4 ε k + ε k ∣ v ~ k ′ ∣ 2 d t − j 2 4 M k 2 . (140) \begin{aligned}
\int_I \left(M_k^2 \wedge\varepsilon_k h_j^2(v_k)\right)|\beta_k'|^2 + \frac{(1-v_k)^2}{4\varepsilon_k}+\varepsilon_k|v_k'|^2 \,\mathrm{d}t
&\ge\int_I \varepsilon_k h_j^2(\widetilde{v}_k)|\beta_k'|^2 + \frac{(1-\widetilde{v}_k)^2}{4\varepsilon_k}+\varepsilon_k|\widetilde{v}_k'|^2 \,\mathrm{d}t \\
&\quad-\int_{\{v_k>\gamma_k\}}\frac{(1-\gamma_k)^2}{4\varepsilon_k}\,\mathrm{d}t \\
&\ge\int_I \varepsilon_k h_j^2(\widetilde{v}_k)|\beta_k'|^2 + \frac{(1-\widetilde{v}_k)^2}{4\varepsilon_k}+\varepsilon_k|\widetilde{v}_k'|^2 \,\mathrm{d}t-\frac{j^2}{4M_k^2}. \tag*{(140)}
\end{aligned} ∫ I ( M k 2 ∧ ε k h j 2 ( v k ) ) ∣ β k ′ ∣ 2 + 4 ε k ( 1 − v k ) 2 + ε k ∣ v k ′ ∣ 2 d t ≥ ∫ I ε k h j 2 ( v k ) ∣ β k ′ ∣ 2 + 4 ε k ( 1 − v k ) 2 + ε k ∣ v k ′ ∣ 2 d t − ∫ { v k > γ k } 4 ε k ( 1 − γ k ) 2 d t ≥ ∫ I ε k h j 2 ( v k ) ∣ β k ′ ∣ 2 + 4 ε k ( 1 − v k ) 2 + ε k ∣ v k ′ ∣ 2 d t − 4 M k 2 j 2 . (140)
Therefore
∫ I ( M k 2 ∧ ε k f 2 ( v k ) ) ∣ β k ′ ∣ 2 d t + PF ε k ( v k ; I ) + 1 δ k ∫ I dist 2 ( β k , O ( d ) ) d t ≥ ∫ I ( 1 − v ~ k ) 2 + 4 ε k δ k dist 2 ( β k , O ( d ) ) h j 2 ( v ~ k ) ∣ β k ′ ∣ 2 + ∣ v ~ k ′ ∣ 2 d t − j 2 4 M k 2 . (141) \begin{aligned}
\int_I \left(M_k^2 \wedge\varepsilon_k f^2(v_k)\right)|\beta_k'|^2 \,\mathrm{d}t+\operatorname{PF}_{\varepsilon_k}(v_k;I)+\frac{1}{\delta_k}\int_I\operatorname{dist}^2(\beta_k,\mathrm{O}(d))\,\mathrm{d}t \\
\ge\int_I \sqrt{(1-\widetilde{v}_k)^2+4\frac{\varepsilon_k}{\delta_k}\operatorname{dist}^2(\beta_k,\mathrm{O}(d))}\sqrt{h_j^2(\widetilde{v}_k)|\beta_k'|^2+|\widetilde{v}_k'|^2}\,\mathrm{d}t-\frac{j^2}{4M_k^2}. \tag*{(141)}
\end{aligned} ∫ I ( M k 2 ∧ ε k f 2 ( v k ) ) ∣ β k ′ ∣ 2 d t + PF ε k ( v k ; I ) + δ k 1 ∫ I dist 2 ( β k , O ( d )) d t ≥ ∫ I ( 1 − v k ) 2 + 4 δ k ε k dist 2 ( β k , O ( d )) h j 2 ( v k ) ∣ β k ′ ∣ 2 + ∣ v k ′ ∣ 2 d t − 4 M k 2 j 2 . (141)
The last term tends to zero for k → ∞ k\to\infty k → ∞ . In order to estimate the integral, we distinguish three cases. If λ = 0 \lambda=0 λ = 0 , the integral on right-hand side of this equation is larger than or equal to H j , 0 ( β k , v ~ k ) H_{j,0}(\beta_k,\widetilde{v}_k) H j , 0 ( β k , v k ) and we conclude using Lemma 5.3. If λ ∈ ( 0 , ∞ ) \lambda\in(0,\infty) λ ∈ ( 0 , ∞ ) , the same integral is larger than or equal to ρ k 1 / 2 H j , λ ( β k , v ~ k ) \rho_k^{1/2}H_{j,\lambda}(\beta_k,\widetilde{v}_k) ρ k 1/2 H j , λ ( β k , v k ) , where ρ k : = 1 ∧ ε k λ δ k → 1 \rho_k:=1\wedge\frac{\varepsilon_k}{\lambda\delta_k}\to1 ρ k := 1 ∧ λ δ k ε k → 1 , and the conclusion follows similarly. Indeed, for any a , b ≥ 0 a,b\ge0 a , b ≥ 0 ,
ρ k 1 / 2 a + λ b = ρ k a + ρ k λ b ≤ a + ε k δ k b . \rho_k^{1/2}\sqrt{a+\lambda b}=\sqrt{\rho_k a+\rho_k\lambda b}\le\sqrt{a+\frac{\varepsilon_k}{\delta_k}b}. ρ k 1/2 a + λb = ρ k a + ρ k λb ≤ a + δ k ε k b .
It remains to treat the case ε k / δ k → ∞ \varepsilon_k/\delta_k\to\infty ε k / δ k → ∞ . Fix λ ′ > 1 \lambda'>1 λ ′ > 1 . For k k k sufficiently large, we have ε k / δ k ≥ λ ′ \varepsilon_k/\delta_k\ge\lambda' ε k / δ k ≥ λ ′ and therefore the integral in (5.23) is larger or equal than H j , λ ′ ( β k , v ~ k ) H_{j,\lambda'}(\beta_k,\widetilde{v}_k) H j , λ ′ ( β k , v k ) . The conclusion now follows by taking the supremum in λ ′ \lambda' λ ′ , in view of (5.3) (and recalling again (5.2)). □ \square □
We are ready to prove the lower bound, by blow-up and slicing.
Proof of Theorem 5.1. If lim inf ε → 0 F ~ ε ( u ε , v ε ) = + ∞ \liminf_{\varepsilon\to0}\widetilde{\mathcal{F}}_\varepsilon(u_\varepsilon,v_\varepsilon)=+\infty lim inf ε → 0 F ε ( u ε , v ε ) = + ∞ we have nothing to prove. On the other hand, if lim inf ε → 0 F ~ ε ( u ε , v ε ) < + ∞ \liminf_{\varepsilon\to0}\widetilde{\mathcal{F}}_\varepsilon(u_\varepsilon,v_\varepsilon)<+\infty lim inf ε → 0 F ε ( u ε , v ε ) < + ∞ by Theorem 3.2 we may assume, possibly passing to a subsequence and with a small abuse of notation, that v = 1 v=1 v = 1 a.e. in Ω \Omega Ω , u ∈ S B V 0 ( Ω ; O ( d ) / G ) u\in\mathrm{SBV}_0(\Omega;\mathrm{O}(d)/\mathcal{G}) u ∈ SBV 0 ( Ω ; O ( d ) / G ) , and
lim inf ε → 0 F ~ ε ( u ε , v ε ) = lim ε → 0 F ~ ε ( u ε , v ε ) . \liminf_{\varepsilon\to0}\widetilde{\mathcal{F}}_\varepsilon(u_\varepsilon,v_\varepsilon)=\lim_{\varepsilon\to0}\widetilde{\mathcal{F}}_\varepsilon(u_\varepsilon,v_\varepsilon). ε → 0 lim inf F ε ( u ε , v ε ) = ε → 0 lim F ε ( u ε , v ε ) .
We define the nonnegative Radon measures
μ ε : = ( f ε 2 ( v ε ) ∣ ∇ u ε ∣ 2 + ( 1 − v ε ) 2 4 ε + ε ∣ ∇ v ε ∣ 2 + 1 δ ε d 2 ( u ε , O ( d ) / G ) ) L N , (142) \mu_\varepsilon:=\left(f_\varepsilon^2(v_\varepsilon)|\nabla u_\varepsilon|^2+\frac{(1-v_\varepsilon)^2}{4\varepsilon}+\varepsilon|\nabla v_\varepsilon|^2+\frac{1}{\delta_\varepsilon}d^2(u_\varepsilon,\mathrm{O}(d)/\mathcal{G})\right)\mathcal{L}^N, \tag*{(142)} μ ε := ( f ε 2 ( v ε ) ∣∇ u ε ∣ 2 + 4 ε ( 1 − v ε ) 2 + ε ∣∇ v ε ∣ 2 + δ ε 1 d 2 ( u ε , O ( d ) / G ) ) L N , (142)
so that F ~ ε ( u ε , v ε ) = μ ε ( Ω ) \widetilde{\mathcal{F}}_\varepsilon(u_\varepsilon,v_\varepsilon)=\mu_\varepsilon(\Omega) F ε ( u ε , v ε ) = μ ε ( Ω ) , and the uniform bound for the energies corresponds to a uniform bound for the total variation of the measures μ ε \mu_\varepsilon μ ε , and hence there exists a nonnegative Radon measure μ \mu μ in Ω \Omega Ω which is the weak*–limit of μ ε \mu_\varepsilon μ ε (up to extracting a further subsequence, not relabeled). By the lower semicontinuity of the total variation with respect to the weak* convergence of measures, we have that
μ ( Ω ) ≤ lim inf ε → 0 μ ε ( Ω ) = lim ε → 0 F ~ ε ( u ε , v ε ) . \mu(\Omega)\le\liminf_{\varepsilon\to0}\mu_\varepsilon(\Omega)=\lim_{\varepsilon\to0}\widetilde{\mathcal{F}}_\varepsilon(u_\varepsilon,v_\varepsilon). μ ( Ω ) ≤ ε → 0 lim inf μ ε ( Ω ) = ε → 0 lim F ε ( u ε , v ε ) .
and therefore the result is proved once we show that
μ ( Ω ) ≥ F ~ ( u , 1 ) , (143) \mu(\Omega) \ge\widetilde{\mathcal{F}}(u,1), \tag*{(143)} μ ( Ω ) ≥ F ( u , 1 ) , (143)
or, equivalently, that for H N − 1 \mathcal{H}^{N-1} H N − 1 -a.e. x 0 ∈ J u x_0 \in J_u x 0 ∈ J u it holds
\frac{\mathrm{d}\mu}{\mathrm{d}(\mathcal{H}^{N-1}\mathbin{\vrule height 1.4ex depth -0.4ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.7ex}J_u)}(x_0) \ge g_\lambda(R^-,R^+) \tag*{(144)}
where R ± ∈ u ± ( x 0 ) R^\pm\in u^\pm(x_0) R ± ∈ u ± ( x 0 ) are any representatives of the equivalence class, the choice being irrelevant by the invariance of g λ g_\lambda g λ under the separate action of G \mathcal{G} G on its two arguments.
In the following, we prove the last inequality for x 0 ∈ J u x_0 \in J_u x 0 ∈ J u such that the limit
\lim_{\rho\to0+}\frac{\mu(Q_\rho^\nu(x_0))}{\mathcal{H}^{N-1}(J_u\cap Q_\rho^\nu(x_0))}=\frac{\mathrm{d}\mu}{\mathrm{d}(\mathcal{H}^{N-1}\mathbin{\vrule height 1.4ex depth -0.4ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.7ex}J_u)}(x_0) \tag*{(145)}
exists finite, and
lim ρ → 0 + H N − 1 ( J u ∩ Q ρ ν ( x 0 ) ) ρ N − 1 = 1 , (146) \lim_{\rho\to0+}\frac{\mathcal{H}^{N-1}(J_u\cap Q_\rho^\nu(x_0))}{\rho^{N-1}}=1, \tag*{(146)} ρ → 0 + lim ρ N − 1 H N − 1 ( J u ∩ Q ρ ν ( x 0 )) = 1 , (146)
where ν : = ν u ( x 0 ) \nu:=\nu_u(x_0) ν := ν u ( x 0 ) , and Q ρ ν ( x 0 ) = x 0 + ρ Q ν Q_\rho^\nu(x_0)=x_0+\rho Q^\nu Q ρ ν ( x 0 ) = x 0 + ρ Q ν is the cube centered in x 0 x_0 x 0 with a face orthogonal to ν \nu ν and side length ρ \rho ρ .
By [1 Theorem 2.3] , J u J_u J u is a H N − 1 \mathcal{H}^{N-1} H N − 1 countably rectifiable set, and hence (5.27) and (5.28) are satisfied H N − 1 \mathcal{H}^{N-1} H N − 1 -a.e. in J u J_u J u . Fixed x 0 x_0 x 0 as above, by weak∗ ^* ∗ -convergence of measures, we have that lim ε → 0 μ ε ( Q ρ ν ( x 0 ) ) = μ ( Q ρ ν ( x 0 ) ) \lim_{\varepsilon\to0}\mu_\varepsilon(Q_\rho^\nu(x_0))=\mu(Q_\rho^\nu(x_0)) lim ε → 0 μ ε ( Q ρ ν ( x 0 )) = μ ( Q ρ ν ( x 0 )) for every ρ \rho ρ such that μ ( ∂ Q ρ ν ( x 0 ) ) = 0 \mu(\partial Q_\rho^\nu(x_0))=0 μ ( ∂ Q ρ ν ( x 0 )) = 0 . Then, by diagonalization, we can find sequences ρ k → 0 \rho_k\to0 ρ k → 0 , ε k → 0 \varepsilon_k\to0 ε k → 0 with ε k / ρ k → 0 \varepsilon_k/\rho_k\to0 ε k / ρ k → 0 , δ ε k / ρ k → 0 \delta_{\varepsilon_k}/\rho_k\to0 δ ε k / ρ k → 0 and such that, setting u k : = u ε k u_k:=u_{\varepsilon_k} u k := u ε k , v k : = v ε k v_k:=v_{\varepsilon_k} v k := v ε k , it holds
\frac{\mathrm{d}\mu}{\mathrm{d}(\mathcal{H}^{N-1}\mathbin{\vrule height 1.4ex depth -0.4ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.7ex}J_u)}(x_0)
=\lim_{k\to+\infty}\frac{\mu_{\varepsilon_k}(Q_{\rho_k}^\nu(x_0))}{\rho_k^{N-1}}
=\lim_{k\to+\infty}\frac{1}{\rho_k^{N-1}}\widetilde{\mathcal{F}}_{\varepsilon_k}(u_k,v_k;Q_{\rho_k}^\nu(x_0))
and
ρ k − N ∥ d G ( u k , [ R ‾ ] ( ⋅ + x 0 ) ) + ∣ v k − 1 ∣ ∥ L 1 ( Q ρ k ν ( x 0 ) ) → 0 , \rho_k^{-N}\left\| \mathrm{d}_{\mathcal{G}}(u_k,[\overline{R}](\cdot+x_0))+|v_k-1|\right\|_{L^1(Q_{\rho_k}^\nu(x_0))}\to0, ρ k − N d G ( u k , [ R ] ( ⋅ + x 0 )) + ∣ v k − 1∣ L 1 ( Q ρ k ν ( x 0 )) → 0 ,
where
R ‾ : = R − χ { x ⋅ ν < 0 } + R + χ { x ⋅ ν > 0 } . \overline{R}:=R^-\chi_{\{x\cdot\nu<0\}}+R^+\chi_{\{x\cdot\nu>0\}}. R := R − χ { x ⋅ ν < 0 } + R + χ { x ⋅ ν > 0 } .
Changing variable to y : = ( x − x 0 ) / ρ k ∈ Q ν y:=(x-x_0)/\rho_k\in Q^\nu y := ( x − x 0 ) / ρ k ∈ Q ν in the integrals defining F ~ ε k ( u k , v k ; Q ρ k ν ( x 0 ) ) \widetilde{\mathcal{F}}_{\varepsilon_k}(u_k,v_k;Q_{\rho_k}^\nu(x_0)) F ε k ( u k , v k ; Q ρ k ν ( x 0 )) , and denoting as usual for the rescaled functions by
u ^ k ( y ) : = u k ( x 0 + ρ k y ) , v ^ k ( y ) : = v k ( x 0 + ρ k y ) , y ∈ Q ν , \widehat{u}_k(y):=u_k(x_0+\rho_k y),\qquad\widehat{v}_k(y):=v_k(x_0+\rho_k y),\qquad y\in Q^\nu, u k ( y ) := u k ( x 0 + ρ k y ) , v k ( y ) := v k ( x 0 + ρ k y ) , y ∈ Q ν ,
we get d G ( u ^ k , [ R ‾ ] ) → 0 \mathrm{d}_{\mathcal{G}}(\widehat{u}_k,[\overline{R}])\to0 d G ( u k , [ R ]) → 0 , v ^ k → 1 \widehat{v}_k\to1 v k → 1 in L 1 ( Q ν ) L^1(Q^\nu) L 1 ( Q ν ) , and
1 ρ k N − 1 F ~ ε k ( u k , v k ; Q ρ k ν ( x 0 ) ) = ∫ Q ν ( ( M ^ k 2 ∧ ε ^ k f 2 ( v ^ k ) ) ∣ ∇ u ^ k ∣ 2 + ( 1 − v ^ k ) 2 4 ε ^ k + ε ^ k ∣ ∇ v ^ k ∣ 2 + 1 δ ^ k d G 2 ( u ^ k , O ( d ) / G ) ) d y \frac{1}{\rho_k^{N-1}}\widetilde{\mathcal{F}}_{\varepsilon_k}(u_k,v_k;Q_{\rho_k}^\nu(x_0))
=\int_{Q^\nu}\left((\widehat{M}_k^2\wedge\widehat{\varepsilon}_k f^2(\widehat{v}_k))|\nabla\widehat{u}_k|^2+\frac{(1-\widehat{v}_k)^2}{4\widehat{\varepsilon}_k}+\widehat{\varepsilon}_k|\nabla\widehat{v}_k|^2+\frac{1}{\widehat{\delta}_k}\mathrm{d}_{\mathcal{G}}^2(\widehat{u}_k,O(d)/\mathcal{G})\right)\,\mathrm{d}y ρ k N − 1 1 F ε k ( u k , v k ; Q ρ k ν ( x 0 )) = ∫ Q ν ( ( M k 2 ∧ ε k f 2 ( v k )) ∣∇ u k ∣ 2 + 4 ε k ( 1 − v k ) 2 + ε k ∣∇ v k ∣ 2 + δ k 1 d G 2 ( u k , O ( d ) / G ) ) d y
for
M ^ k : = M ε k ρ k , ε ^ k : = ε k ρ k , δ ^ k : = δ ε k ρ k , so that lim k → + ∞ ε ^ k δ ^ k = lim k → + ∞ ε k δ ε k = λ . \widehat{M}_k:=\frac{M_{\varepsilon_k}}{\sqrt{\rho_k}},\qquad\widehat{\varepsilon}_k:=\frac{\varepsilon_k}{\rho_k},\qquad\widehat{\delta}_k:=\frac{\delta_{\varepsilon_k}}{\rho_k},\qquad\text{so that}\qquad\lim_{k\to+\infty}\frac{\widehat{\varepsilon}_k}{\widehat{\delta}_k}=\lim_{k\to+\infty}\frac{\varepsilon_k}{\delta_{\varepsilon_k}}=\lambda. M k := ρ k M ε k , ε k := ρ k ε k , δ k := ρ k δ ε k , so that k → + ∞ lim δ k ε k = k → + ∞ lim δ ε k ε k = λ .
Hence, if we consider the slices of these functions in the direction ν \nu ν , defined for z ∈ Π ν ∩ Q ν : = { z ∈ Q ν : z ⋅ ν = 0 } z\in\Pi^\nu\cap Q^\nu:=\{z\in Q^\nu:z\cdot\nu=0\} z ∈ Π ν ∩ Q ν := { z ∈ Q ν : z ⋅ ν = 0 } as
( u ^ k ) z ν ( t ) : = u ^ k ( z + t ν ) , ( v ^ k ) z ν ( t ) : = v ^ k ( z + t ν ) , t ∈ I = [ − 1 2 , 1 2 ] , (\widehat{u}_k)_z^\nu(t):=\widehat{u}_k(z+t\nu),\qquad(\widehat{v}_k)_z^\nu(t):=\widehat{v}_k(z+t\nu),\qquad t\in I=\left[-\frac{1}{2},\frac{1}{2}\right], ( u k ) z ν ( t ) := u k ( z + t ν ) , ( v k ) z ν ( t ) := v k ( z + t ν ) , t ∈ I = [ − 2 1 , 2 1 ] ,
then, by Proposition 2.11, for H N − 1 \mathcal{H}^{N-1} H N − 1 –a.e. z ∈ Π ν ∩ Q ν z \in\Pi^\nu\cap Q^\nu z ∈ Π ν ∩ Q ν ,
( ( u ^ k ) z ν , ( v ^ k ) z ν ) ∈ H 1 ( I ; M d × d / G × [ 0 , 1 ] ) , (147) \left((\widehat{u}_k)^\nu_z,(\widehat{v}_k)^\nu_z\right) \in H^1(I;M^{d\times d}/\mathcal{G}\times[0,1]), \tag*{(147)} ( ( u k ) z ν , ( v k ) z ν ) ∈ H 1 ( I ; M d × d / G × [ 0 , 1 ]) , (147)
and
( u ^ k ) z ν → [ R − ] χ [ − 1 / 2 , 0 ) + [ R + ] χ [ 0 , 1 / 2 ] in L 1 ( I ; M d × d / G ) , ( v ^ k ) z ν → 1 in L 1 ( I ) . (148) \begin{aligned}
(\widehat{u}_k)^\nu_z &\to[R^-]\chi_{[-1/2,0)}+[R^+]\chi_{[0,1/2]} \quad\text{in }L^1(I;M^{d\times d}/\mathcal{G}),\\
(\widehat{v}_k)^\nu_z &\to1 \quad\text{in }L^1(I). \tag*{(148)}
\end{aligned} ( u k ) z ν ( v k ) z ν → [ R − ] χ [ − 1/2 , 0 ) + [ R + ] χ [ 0 , 1/2 ] in L 1 ( I ; M d × d / G ) , → 1 in L 1 ( I ) . (148)
By Lemma 2.5, for every k ∈ N k\in\mathbb{N} k ∈ N , and for every z ∈ Π ν ∩ Q ν z\in\Pi^\nu\cap Q^\nu z ∈ Π ν ∩ Q ν such that (5.29) holds true, there exists β k , z ∈ H 1 ( I ; M d × d ) \beta_{k,z}\in H^1(I;M^{d\times d}) β k , z ∈ H 1 ( I ; M d × d ) such that [ β k , z ] = ( u ^ k ) z ν [\beta_{k,z}]=(\widehat{u}_k)^\nu_z [ β k , z ] = ( u k ) z ν and ∣ β k , z ′ ∣ = ∣ ( ( u ^ k ) z ν ) ′ ∣ \lvert\beta'_{k,z}\rvert=\lvert((\widehat{u}_k)^\nu_z)'\rvert ∣ β k , z ′ ∣ = ∣(( u k ) z ν ) ′ ∣ , and hence
F ^ k ( ( u ^ k ) z ν , ( v ^ k ) z ν ; I ) : = ∫ I ( M ^ k 2 ∧ ε ^ k f 2 ( ( v ^ k ) z ν ) ) ∣ ( ( u ^ k ) z ν ) ′ ∣ 2 + ( 1 − ( v ^ k ) z ν ) 2 4 ε ^ k 2 + ε ^ k ∣ ( ( v ^ k ) z ν ) ′ ∣ 2 + 1 δ ^ k d G 2 ( ( u ^ k ) z ν , O ( d ) / G ) d t = ∫ I ( M ^ k 2 ∧ ε ^ k f 2 ( ( v ^ k ) z ν ) ) ∣ β k , z ′ ∣ 2 + ( 1 − ( v ^ k ) z ν ) 2 4 ε ^ k 2 + ε ^ k ∣ ( ( v ^ k ) z ν ) ′ ∣ 2 + 1 δ ^ k dist 2 ( β k , z , O ( d ) ) d t . \begin{aligned}
\widehat{F}_k((\widehat{u}_k)^\nu_z,(\widehat{v}_k)^\nu_z;I)
&:=\int_I\left(\widehat{M}_k^2\wedge\widehat{\varepsilon}_k f^2((\widehat{v}_k)^\nu_z)\right)\lvert((\widehat{u}_k)^\nu_z)'\rvert^2+\frac{(1-(\widehat{v}_k)^\nu_z)^2}{4\widehat{\varepsilon}_k^2}+\widehat{\varepsilon}_k\lvert((\widehat{v}_k)^\nu_z)'\rvert^2+\frac{1}{\widehat{\delta}_k}d_{\mathcal{G}}^2((\widehat{u}_k)^\nu_z,O(d)/\mathcal{G})\,dt\\
&=\int_I\left(\widehat{M}_k^2\wedge\widehat{\varepsilon}_k f^2((\widehat{v}_k)^\nu_z)\right)\lvert\beta'_{k,z}\rvert^2+\frac{(1-(\widehat{v}_k)^\nu_z)^2}{4\widehat{\varepsilon}_k^2}+\widehat{\varepsilon}_k\lvert((\widehat{v}_k)^\nu_z)'\rvert^2+\frac{1}{\widehat{\delta}_k}\operatorname{dist}^2(\beta_{k,z},O(d))\,dt.
\end{aligned} F k (( u k ) z ν , ( v k ) z ν ; I ) := ∫ I ( M k 2 ∧ ε k f 2 (( v k ) z ν ) ) ∣(( u k ) z ν ) ′ ∣ 2 + 4 ε k 2 ( 1 − ( v k ) z ν ) 2 + ε k ∣(( v k ) z ν ) ′ ∣ 2 + δ k 1 d G 2 (( u k ) z ν , O ( d ) / G ) d t = ∫ I ( M k 2 ∧ ε k f 2 (( v k ) z ν ) ) ∣ β k , z ′ ∣ 2 + 4 ε k 2 ( 1 − ( v k ) z ν ) 2 + ε k ∣(( v k ) z ν ) ′ ∣ 2 + δ k 1 dist 2 ( β k , z , O ( d )) d t .
Moreover, by (148) ,
β k , z → L G 1 R − χ [ − 1 / 2 , 0 ) + R + χ [ 0 , 1 / 2 ] , for H N − 1 –a.e. z ∈ Π ν ∩ Q ν . \beta_{k,z}\xrightarrow[L^1_{\mathcal{G}}]{}R^-\chi_{[-1/2,0)}+R^+\chi_{[0,1/2]},\qquad\text{for }\mathcal{H}^{N-1}\text{–a.e. }z\in\Pi^\nu\cap Q^\nu. β k , z L G 1 R − χ [ − 1/2 , 0 ) + R + χ [ 0 , 1/2 ] , for H N − 1 –a.e. z ∈ Π ν ∩ Q ν .
By Lemma 5.2 and Lemma 5.4 , we obtain
lim inf k → + ∞ F ^ k ( ( u ^ k ) z ν , ( v ^ k ) z ν ; I ) ≥ g λ ( R − , R + ) , \liminf_{k\to+\infty}\widehat{F}_k((\widehat{u}_k)^\nu_z,(\widehat{v}_k)^\nu_z;I)\ge g_\lambda(R^-,R^+), k → + ∞ lim inf F k (( u k ) z ν , ( v k ) z ν ; I ) ≥ g λ ( R − , R + ) ,
for H N − 1 \mathcal{H}^{N-1} H N − 1 –a.e. z ∈ Π ν ∩ Q ν z\in\Pi^\nu\cap Q^\nu z ∈ Π ν ∩ Q ν .
The validity of (144) , and hence the conclusion, now follows by integrating on Π ν \Pi^\nu Π ν , and using Fubini’s theorem, the estimates in Proposition 2.11, and Fatou’s lemma:
\frac{d\mu}{d(\mathcal{H}^{N-1}\mathbin{\vrule height 1.4ex depth -0.4ex width 0.07ex\vrule height 0.07ex depth -0.02ex width 0.7ex}J_u)}(x_0)
=\lim_{k\to+\infty}\frac{1}{\rho_k^{N-1}}\widetilde{F}_{\varepsilon_k}(u_k,v_k;Q_{\rho_k}(x_0))
\ge\lim_{k\to+\infty}\int_{\Pi^\nu\cap Q^\nu}\widehat{F}_k((\widehat{u}_k)^\nu_z,(\widehat{v}_k)^\nu_z;I)\,dz
\ge\int_{\Pi^\nu\cap Q^\nu}\liminf_{k\to+\infty}\widehat{F}_k((\widehat{u}_k)^\nu_z,(\widehat{v}_k)^\nu_z;I)\,dz
\ge g_\lambda(R^-,R^+).
Remark 5.5. We have proven above (LB)’ in Theorem 2.15. The analogous (LB) in Theorem 1.1 can be obtained either directly, following the same proof proposed in Theorem 5.1, or as a consequence of Remark 2.1, (2.31), (2.32), and Theorem 5.1.
Upper bound
In this section we prove the upper bound for the functionals F ε F_\varepsilon F ε stated in Theorem 1.1(UB).
Theorem 5.6. For any β ∈ S B V 0 ( Ω ; O ( d ) ) \beta\in SBV_0(\Omega;O(d)) β ∈ S B V 0 ( Ω ; O ( d )) there exists ( β ε , v ε ) ε ⊂ S B V G ( Ω ; M d × d ) × H 1 ( Ω ) (\beta_\varepsilon,v_\varepsilon)_\varepsilon\subset SBV_{\mathcal{G}}(\Omega;M^{d\times d})\times H^1(\Omega) ( β ε , v ε ) ε ⊂ S B V G ( Ω ; M d × d ) × H 1 ( Ω ) , with
β ε → β in L G 1 ( Ω ; M d × d ) , v ε → 1 in L 1 ( Ω ) , \beta_\varepsilon\to\beta\quad\text{in }L^1_{\mathcal{G}}(\Omega;M^{d\times d}),\qquad v_\varepsilon\to1\quad\text{in }L^1(\Omega), β ε → β in L G 1 ( Ω ; M d × d ) , v ε → 1 in L 1 ( Ω ) , and such that
F λ ( β , 1 ) ≥ lim sup ε → 0 F ε ( β ε , v ε ) . (149) \mathcal{F}^\lambda(\beta,1)\ge\limsup_{\varepsilon\to0}\mathcal{F}_\varepsilon(\beta_\varepsilon,v_\varepsilon). \tag*{(149)} F λ ( β , 1 ) ≥ ε → 0 lim sup F ε ( β ε , v ε ) . (149) Moreover, for any sequence η ε → 0 \eta_\varepsilon\to0 η ε → 0 such that
lim ε → 0 η ε ε = 0 and lim ε → 0 η ε δ ε = 0 (150) \lim_{\varepsilon\to0}\frac{\eta_\varepsilon}{\varepsilon}=0\qquad\text{and}\qquad\lim_{\varepsilon\to0}\frac{\eta_\varepsilon}{\delta_\varepsilon}=0 \tag*{(150)} ε → 0 lim ε η ε = 0 and ε → 0 lim δ ε η ε = 0 (150) one can choose the sequences ( β ε , v ε ) (\beta_\varepsilon,v_\varepsilon) ( β ε , v ε ) such that η ε ∥ ∇ β ε ∥ L 2 ( Ω ) 2 → 0 \eta_\varepsilon\lVert\nabla\beta_\varepsilon\rVert_{L^2(\Omega)}^2\to0 η ε ∥ ∇ β ε ∥ L 2 ( Ω ) 2 → 0 .
If λ < ∞ \lambda< \infty λ < ∞ then it suffices to consider the first condition in (150) , if λ = ∞ \lambda= \infty λ = ∞ it suffices to consider the second one.
Proof of Theorem 5.6. We first remark that the density result [7 Lemma 4.1] , valid for sharp-interface energies depending on the jump, holds also for our energies F λ \mathcal{F}^{\lambda} F λ that depend on the two traces. Namely, with the very same proof of [7 Lemma 4.1] one obtains that for every given β ∈ S B V 0 ( Ω ; O ( d ) ) \beta\in\mathrm{SBV}_{0}(\Omega;\mathrm{O}(d)) β ∈ SBV 0 ( Ω ; O ( d )) there exist β k ∈ S B V 0 ( Ω k ; O ( d ) ) \beta_k \in\mathrm{SBV}_{0}(\Omega_k;\mathrm{O}(d)) β k ∈ SBV 0 ( Ω k ; O ( d )) such that
Ω ⋐ Ω k β k → β in L 1 ( Ω ; O ( d ) ) , F λ ( β , 1 ) = lim k → + ∞ F λ ( β k , 1 ; Ω k ) , \Omega\Subset\Omega_k \qquad\beta_k \to\beta\quad\text{in } L^1(\Omega;\mathrm{O}(d)), \qquad\mathcal{F}^{\lambda}(\beta,1)=\lim_{k\to+\infty}\mathcal{F}^{\lambda}(\beta_k,1;\Omega_k), Ω ⋐ Ω k β k → β in L 1 ( Ω ; O ( d )) , F λ ( β , 1 ) = k → + ∞ lim F λ ( β k , 1 ; Ω k ) ,
J β k J_{\beta_k} J β k is a finite union simplexes of dimension N − 1 N-1 N − 1 .
The map β k \beta_k β k constructed in [7 Lemma 4.1] takes values in O ( d ) \mathrm{O}(d) O ( d ) because it coincides in any point of Ω \Omega Ω with a suitable value of β \beta β .
Therefore we can assume that Ω ⋐ Ω ′ \Omega\Subset\Omega' Ω ⋐ Ω ′ , β ∈ S B V 0 ( Ω ′ ; O ( d ) ) \beta\in\mathrm{SBV}_{0}(\Omega';\mathrm{O}(d)) β ∈ SBV 0 ( Ω ′ ; O ( d )) is piecewise constant on finitely many convex polyhedra, whose faces are unions of simplexes of dimension N − 1 N-1 N − 1 , which are denoted by F l F_l F l . We need to find sequences ( β ε , v ε ) → ( β , 1 ) (\beta_\varepsilon,v_\varepsilon)\to(\beta,1) ( β ε , v ε ) → ( β , 1 ) in L 1 ( Ω ) L^1(\Omega) L 1 ( Ω ) such that
lim sup ε → 0 F ε ( β ε , v ε ) ≤ F λ ( β , 1 ; Ω ′ ) = ∑ l H N − 1 ( F l ) g λ ( R l − , R l + ) , \limsup_{\varepsilon\to0}\mathcal{F}_\varepsilon(\beta_\varepsilon,v_\varepsilon)\leq\mathcal{F}^{\lambda}(\beta,1;\Omega')=\sum_l\mathcal{H}^{N-1}(F_l)g_\lambda(R_l^-,R_l^+), ε → 0 lim sup F ε ( β ε , v ε ) ≤ F λ ( β , 1 ; Ω ′ ) = l ∑ H N − 1 ( F l ) g λ ( R l − , R l + ) ,
where R l ± ∈ O ( d ) R_l^\pm\in\mathrm{O}(d) R l ± ∈ O ( d ) are the two traces of β \beta β on the face F l F_l F l . We start with a decomposition of the domain, referring to Figure 2 for an illustration. We denote by ν l \nu_l ν l the normal to F l F_l F l and by ∂ ′ F l \partial'F_l ∂ ′ F l the relative boundary of F l F_l F l , in the sense of the boundary of F l F_l F l seen as a subset of a hyperplane. Possibly subdividing some of the facets, we can ensure that they have disjoint ( N − 1 ) (N-1) ( N − 1 ) -dimensional interior, in the sense that the sets F l ∖ ∂ ′ F l F_l\setminus\partial'F_l F l ∖ ∂ ′ F l are disjoint. For y ∈ F l y\in F_l y ∈ F l we write d l ( y ) d_l(y) d l ( y ) for the distance to the boundary, d l ( y ) : = dist ( y , ∂ ′ F l ) d_l(y):=\operatorname{dist}(y,\partial'F_l) d l ( y ) := dist ( y , ∂ ′ F l ) , and write briefly Γ : = ⋃ l ∂ ′ F l \Gamma:=\bigcup_l\partial'F_l Γ := ⋃ l ∂ ′ F l .
Figure 2. Sketch of the construction in the upper bound. The black lines represent the facets F l F_l F l , and the orange subset is E ε l E_\varepsilon^l E ε l . The dashed green circles with radius 2 M ε 2M\varepsilon 2 M ε around ∂ ′ F l \partial'F_l ∂ ′ F l represent the set { v ε 0 = 0 } \{v_\varepsilon^0=0\} { v ε 0 = 0 } , the two red ones with radius M ε ± ρ ε M\varepsilon\pm\rho_\varepsilon M ε ± ρ ε denote the area where the two interpolations in (165) take place. The right panel shows a blow-up of a part of the set with some labels.
In what follows we denote by ( E ) r (E)_r ( E ) r the r r r -neighborhood of a set E ⊂ R N E\subset\mathbb{R}^{N} E ⊂ R N . Since each ∂ ′ F l \partial'F_l ∂ ′ F l is a finite union of ( N − 2 ) (N-2) ( N − 2 ) -dimensional polyhedra, we have (see, e.g., [3 Sect. 2.13] )
lim r → 0 ∣ ( ∂ ′ F l ) r ∣ π r 2 = H N − 2 ( ∂ ′ F l ) < ∞ , (151) \lim_{r\to0}\frac{|(\partial'F_l)_r|}{\pi r^2}=\mathcal{H}^{N-2}(\partial'F_l)<\infty, \tag*{(151)} r → 0 lim π r 2 ∣ ( ∂ ′ F l ) r ∣ = H N − 2 ( ∂ ′ F l ) < ∞ , (151)
so that there is C C C with
∣ ( Γ ) r ∣ ≤ C r 2 for all r ∈ ( 0 , 1 ) . (152) |(\Gamma)_r|\leq Cr^2 \quad\text{for all }r\in(0,1). \tag*{(152)} ∣ ( Γ ) r ∣ ≤ C r 2 for all r ∈ ( 0 , 1 ) . (152)
Fix L ≥ 1 L\geq1 L ≥ 1 , we shall in the end take ε → 0 \varepsilon\to0 ε → 0 , L → ∞ L\to\infty L → ∞ , with L ε → 0 L\varepsilon\to0 L ε → 0 and other conditions that will become clear later. We observe that, since the F l F_l F l are ( N − 1 ) (N-1) ( N − 1 ) -dimensional simplexes with disjoint interior, there is c ∗ ∈ [ 1 , ∞ ) c_*\in[1,\infty) c ∗ ∈ [ 1 , ∞ ) such that the sets
E ε l : = { y + t ν l : y ∈ F l , d l ( y ) > c ∗ L ε , t ∈ ( − L ε , L ε ) } (153) E_\varepsilon^l:=\{y+t\nu_l:y\in F_l,d_l(y)>c_*L\varepsilon,\ t\in(-L\varepsilon,L\varepsilon)\} \tag*{(153)} E ε l := { y + t ν l : y ∈ F l , d l ( y ) > c ∗ L ε , t ∈ ( − L ε , L ε )} (153)
are pairwise disjoint if L ε L\varepsilon L ε is sufficiently small, see Figure 2 . In these regions we can use the optimal profiles; we shall than need to interpolate around Γ \Gamma Γ .
We fix M : = ( 1 + c ∗ ) L M:=(1+c_*)L M := ( 1 + c ∗ ) L , so that the lateral boundaries of each E ε l E_\varepsilon^l E ε l are contained in ( ∂ ′ F l ) M ε (\partial'F_l)_{M\varepsilon} ( ∂ ′ F l ) M ε , in the sense that
∂ E ε l ⊂ ( F l + L ε ν l ) ∪ ( F l − L ε ν l ) ∪ ( ∂ ′ F l ) M ε (154) \partial E_\varepsilon^l\subset(F_l+L\varepsilon\nu_l)\cup(F_l-L\varepsilon\nu_l)\cup(\partial'F_l)_{M\varepsilon} \tag*{(154)} ∂ E ε l ⊂ ( F l + L ε ν l ) ∪ ( F l − L ε ν l ) ∪ ( ∂ ′ F l ) M ε (154)
(the relevance of this condition will become clear after (159) below) and define v ε 0 ∈ H 1 ( Ω ; [ 0 , 1 ] ) v_\varepsilon^0\in H^1(\Omega;[0,1]) v ε 0 ∈ H 1 ( Ω ; [ 0 , 1 ]) by
v ε 0 ( x ) : = 1 ∧ 1 M ε dist ( x , ( Γ ) 2 M ε ) . v_\varepsilon^0(x):=1\wedge\frac{1}{M\varepsilon}\operatorname{dist}(x,(\Gamma)_{2M\varepsilon}). v ε 0 ( x ) := 1 ∧ M ε 1 dist ( x , ( Γ ) 2 M ε ) .
We observe that ∥ ∇ v ε 0 ∥ L ∞ ≤ 1 / ( M ε ) \lVert\nabla v_\varepsilon^0\rVert_{L^\infty}\leq1/(M\varepsilon) ∥ ∇ v ε 0 ∥ L ∞ ≤ 1/ ( M ε ) and that v ε 0 = 1 v_\varepsilon^0=1 v ε 0 = 1 outside ( Γ ) 3 M ε (\Gamma)_{3M\varepsilon} ( Γ ) 3 M ε , so that (5.34) implies
P F ε ( v ε 0 ) ≤ ∣ ( Γ ) 3 M ε ∣ ( 1 ε + ε ( M ε ) 2 ) ≤ C L 2 ε . (155) P\mathcal{F}_\varepsilon(v_\varepsilon^0)\leq|(\Gamma)_{3M\varepsilon}|\left(\frac{1}{\varepsilon}+\frac{\varepsilon}{(M\varepsilon)^2}\right)\leq CL^2\varepsilon. \tag*{(155)} P F ε ( v ε 0 ) ≤ ∣ ( Γ ) 3 M ε ∣ ( ε 1 + ( M ε ) 2 ε ) ≤ C L 2 ε . (155)
For each l l l we pick G l ∈ G G_l\in\mathcal{G} G l ∈ G such that g λ ( R l − , R l + ) = g λ ∗ ( R l − , G l − 1 R l + ) g_\lambda(R_l^-,R_l^+)=g_\lambda^*(R_l^-,G_l^{-1}R_l^+) g λ ( R l − , R l + ) = g λ ∗ ( R l − , G l − 1 R l + ) (see (84) and (85) ). If λ < ∞ \lambda<\infty λ < ∞ , by (82) we can select ( β ~ l , v ~ l ) ∈ H 1 ( ( − L , L ) ; M d × d × [ 0 , 1 ] ) (\widetilde{\beta}_l,\widetilde{v}_l)\in H^1((-L,L);M^{d\times d}\times[0,1]) ( β l , v l ) ∈ H 1 (( − L , L ) ; M d × d × [ 0 , 1 ]) such that
v ~ l ( ± L ) = 1 , β ~ l ( − L ) = R l − , G l β ~ l ( L ) = R l + , ∥ β ~ l ∥ L ∞ ≤ d , \widetilde{v}_l(\pm L)=1,\quad\widetilde{\beta}_l(-L)=R_l^-,\quad G_l\widetilde{\beta}_l(L)=R_l^+,\quad\|\widetilde{\beta}_l\|_{L^\infty}\le\sqrt{d}, v l ( ± L ) = 1 , β l ( − L ) = R l − , G l β l ( L ) = R l + , ∥ β l ∥ L ∞ ≤ d ,
∥ β ~ l ′ ∥ L 2 ( ( − L , L ) ) ≤ C ( L ) , F λ L ( β ~ l , v ~ l ) ≤ g λ ( R l + , R l − ) + ω ( L ) , (156) \|\widetilde{\beta}'_l\|_{L^2((-L,L))}\le C(L),\quad\mathcal{F}_\lambda^L(\widetilde{\beta}_l,\widetilde{v}_l)\le g_\lambda(R_l^+,R_l^-)+\omega(L), \tag*{(156)} ∥ β l ′ ∥ L 2 (( − L , L )) ≤ C ( L ) , F λ L ( β l , v l ) ≤ g λ ( R l + , R l − ) + ω ( L ) , (156)
with ω ( L ) → 0 \omega(L)\to0 ω ( L ) → 0 as L → ∞ L\to\infty L → ∞ , and C ( L ) C(L) C ( L ) a positive constant depending only on L L L . If instead λ = ∞ \lambda=\infty λ = ∞ , we use Lemma 4.5(iii) and for any fixed ε > 0 \varepsilon>0 ε > 0 and L > 1 L>1 L > 1 we take
λ ε : = ε δ ε (157) \lambda_\varepsilon:=\frac{\varepsilon}{\delta_\varepsilon} \tag*{(157)} λ ε := δ ε ε (157)
and select ( β ~ l , v ~ l ) ∈ H 1 ( ( − L , L ) ; M d × d × [ 0 , 1 ] ) (\widetilde{\beta}_l,\widetilde{v}_l)\in H^1((-L,L);M^{d\times d}\times[0,1]) ( β l , v l ) ∈ H 1 (( − L , L ) ; M d × d × [ 0 , 1 ]) , depending on ε \varepsilon ε and L L L , so that the first line of (156) still holds, and the second one is replaced by
∥ β ~ l ′ ∥ L 2 ( ( − L , L ) ) ≤ C λ ε L , F λ ε L ( β ~ l , v ~ l ) ≤ g ∞ ( R l + , R l − ) + ω ( L ) + ω ( λ ε ) , (158) \|\widetilde{\beta}'_l\|_{L^2((-L,L))}\le C\lambda_\varepsilon L,\quad\mathcal{F}_{\lambda_\varepsilon}^L(\widetilde{\beta}_l,\widetilde{v}_l)\le g_\infty(R_l^+,R_l^-)+\omega(L)+\omega(\lambda_\varepsilon), \tag*{(158)} ∥ β l ′ ∥ L 2 (( − L , L )) ≤ C λ ε L , F λ ε L ( β l , v l ) ≤ g ∞ ( R l + , R l − ) + ω ( L ) + ω ( λ ε ) , (158)
with ω ( s ) → 0 \omega(s)\to0 ω ( s ) → 0 as s → + ∞ s\to+\infty s → + ∞ . We define v ε l : E ε l → [ 0 , 1 ] v_\varepsilon^l:E_\varepsilon^l\to[0,1] v ε l : E ε l → [ 0 , 1 ] , using y y y and t t t as in (153) , by
v ε l ( y + t ν l ) : = v ~ l ( t / ε ) , (159) v_\varepsilon^l(y+t\nu_l):=\widetilde{v}_l(t/\varepsilon), \tag*{(159)} v ε l ( y + t ν l ) := v l ( t / ε ) , (159)
and then v ε v_\varepsilon v ε by
v ε ( x ) : = { v ε 0 ( x ) ∧ v ε l ( x ) , if x ∈ E ε l , v ε 0 ( x ) , if x ∈ Ω ∖ ⋃ l E ε l . (160) v_\varepsilon(x):=
\begin{cases}
v_\varepsilon^0(x)\wedge v_\varepsilon^l(x), & \text{if }x\in E_\varepsilon^l,\\
v_\varepsilon^0(x), & \text{if }x\in\Omega\setminus\bigcup_l E_\varepsilon^l.
\end{cases} \tag*{(160)} v ε ( x ) := { v ε 0 ( x ) ∧ v ε l ( x ) , v ε 0 ( x ) , if x ∈ E ε l , if x ∈ Ω ∖ ⋃ l E ε l . (160)
We next show that no jumps are introduced on the boundaries of any E ε l E_\varepsilon^l E ε l . Indeed, let x = y + t ν l ∈ ∂ E ε l x=y+t\nu_l\in\partial E_\varepsilon^l x = y + t ν l ∈ ∂ E ε l . If d l ( y ) > c ∗ L ε d_l(y)>c_*L\varepsilon d l ( y ) > c ∗ L ε then ∣ t ∣ = L ε |t|=L\varepsilon ∣ t ∣ = L ε , which implies v ε l ( x ) = v ~ l ( t / ε ) = 1 v_\varepsilon^l(x)=\widetilde{v}_l(t/\varepsilon)=1 v ε l ( x ) = v l ( t / ε ) = 1 . If instead d l ( y ) < c ∗ L ε d_l(y)<c_*L\varepsilon d l ( y ) < c ∗ L ε then dist ( x , ∂ ′ F l ) ≤ L ( 1 + c ∗ ) ε = M ε \operatorname{dist}(x,\partial'F_l)\le L(1+c_*)\varepsilon=M\varepsilon dist ( x , ∂ ′ F l ) ≤ L ( 1 + c ∗ ) ε = M ε , so that v ε 0 ( x ) = 0 v_\varepsilon^0(x)=0 v ε 0 ( x ) = 0 . In both cases, the minimum in the first line of (160) equals v ε 0 ( x ) v_\varepsilon^0(x) v ε 0 ( x ) , and we obtain v ε ∈ H 1 ( Ω ; [ 0 , 1 ] ) v_\varepsilon\in H^1(\Omega;[0,1]) v ε ∈ H 1 ( Ω ; [ 0 , 1 ]) . The monotonicity condition P F ε ( v ′ ∧ v ′ ′ ; A ) ≤ P F ε ( v ′ ; A ) + P F ε ( v ′ ′ ; A ) \mathcal{P}_{F_\varepsilon}(v'\wedge v'';A)\le\mathcal{P}_{F_\varepsilon}(v';A)+\mathcal{P}_{F_\varepsilon}(v'';A) P F ε ( v ′ ∧ v ′′ ; A ) ≤ P F ε ( v ′ ; A ) + P F ε ( v ′′ ; A ) leads directly to
P F ε ( v ε ) ≤ P F ε ( v ε 0 ) + ∑ l P F ε ( v ε l ; E ε l ) . (161) \mathcal{P}_{F_\varepsilon}(v_\varepsilon)\le\mathcal{P}_{F_\varepsilon}(v_\varepsilon^0)+\sum_l\mathcal{P}_{F_\varepsilon}(v_\varepsilon^l;E_\varepsilon^l). \tag*{(161)} P F ε ( v ε ) ≤ P F ε ( v ε 0 ) + l ∑ P F ε ( v ε l ; E ε l ) . (161)
We now turn to the construction of β ε \beta_{\varepsilon} β ε , which is more subtle. To make the strategy clear we deal with each difficulty separately, defining a sequence of intermediate functions.
In order to rescale and account for the matrices G l G_l G l , we replace the functions β ~ l \widetilde{\beta}_l β l identified in (5.38) by
β ε l ( y + t ν l ) : = { G l β ~ l ( t ε ) , if t ∈ [ 0 , L ε ] , β ~ l ( t ε ) , if t ∈ [ − L ε , 0 ) . (162) \beta_{\varepsilon}^{l}(y+t\nu_l):=
\begin{cases}
G_l\widetilde{\beta}_l\left(\frac{t}{\varepsilon}\right), & \text{if }t\in[0,L\varepsilon],\\
\widetilde{\beta}_l\left(\frac{t}{\varepsilon}\right), & \text{if }t\in[-L\varepsilon,0).
\end{cases} \tag*{(162)} β ε l ( y + t ν l ) := { G l β l ( ε t ) , β l ( ε t ) , if t ∈ [ 0 , L ε ] , if t ∈ [ − L ε , 0 ) . (162)
We can easily use this profile in the inner regions E ε l E_{\varepsilon}^{l} E ε l . It automatically matches R l ± R_l^{\pm} R l ± for t = ± L ε t=\pm L\varepsilon t = ± L ε , it can jump only in 0 0 0 , and since β ~ l \widetilde{\beta}_l β l is continuous the traces differ by a factor G l G_l G l . The critical point is the interpolation at the lateral boundaries which are contained in ( ∂ ′ F l ) M ε (\partial'F_l)_{M\varepsilon} ( ∂ ′ F l ) M ε . The condition on the jump required by SBV G \operatorname{SBV}_G SBV G is not linear (in the sense that the sum of two SBV G \operatorname{SBV}_G SBV G functions is not necessarily in SBV G \operatorname{SBV}_G SBV G ), therefore we cannot interpolate. However, SBV G \operatorname{SBV}_G SBV G is one-homogeneous, in the sense that if f ∈ SBV G ( Ω ; M d × d ) f\in\operatorname{SBV}_G(\Omega;\mathbb{M}^{d\times d}) f ∈ SBV G ( Ω ; M d × d ) and ψ ∈ C 1 ( Ω ‾ ) \psi\in C^1(\overline{\Omega}) ψ ∈ C 1 ( Ω ) then ψ f ∈ SBV G ( Ω ; M d × d ) \psi f\in\operatorname{SBV}_G(\Omega;\mathbb{M}^{d\times d}) ψ f ∈ SBV G ( Ω ; M d × d ) . On the other hand, the term δ ε − 1 dist 2 ( ⋅ , O ( d ) ) \delta_{\varepsilon}^{-1}\operatorname{dist}^2(\cdot,O(d)) δ ε − 1 dist 2 ( ⋅ , O ( d )) in the energy may be significantly increased by this operation, therefore we must make sure that { ψ ≠ 1 } \{\psi\ne1\} { ψ = 1 } is small.
Since v ε 0 = 0 v_{\varepsilon}^{0}=0 v ε 0 = 0 in ( Γ ) 2 M ε (\Gamma)_{2M\varepsilon} ( Γ ) 2 M ε , we modify the functions β ε l \beta_{\varepsilon}^{l} β ε l by interpolating first with 0 0 0 and then with the identity. We fix φ ∈ C ∞ ( R ; [ 0 , 1 ] ) \varphi\in C^{\infty}(\mathbb{R};[0,1]) φ ∈ C ∞ ( R ; [ 0 , 1 ]) , such that φ ( t ) = 0 \varphi(t)=0 φ ( t ) = 0 for t < 0 t<0 t < 0 and φ ( t ) = 1 \varphi(t)=1 φ ( t ) = 1 for t ≥ 1 t\ge1 t ≥ 1 , let ρ ε ∈ ( 0 , ε ] \rho_{\varepsilon}\in(0,\varepsilon] ρ ε ∈ ( 0 , ε ] be a parameter to be chosen below, and set
ψ ε o u t ( x ) : = φ ( dist ( x , Γ ) − M ε ρ ε ) ψ ε i n ( x ) : = φ ( M ε − dist ( x , Γ ) ρ ε ) , (163) \psi_{\varepsilon}^{\mathrm{out}}(x):=\varphi\left(\frac{\operatorname{dist}(x,\Gamma)-M\varepsilon}{\rho_{\varepsilon}}\right)
\qquad
\psi_{\varepsilon}^{\mathrm{in}}(x):=\varphi\left(\frac{M\varepsilon-\operatorname{dist}(x,\Gamma)}{\rho_{\varepsilon}}\right), \tag*{(163)} ψ ε out ( x ) := φ ( ρ ε dist ( x , Γ ) − M ε ) ψ ε in ( x ) := φ ( ρ ε M ε − dist ( x , Γ ) ) , (163)
We then define
β ^ ε ( x ) : = { β ε l ( x ) , if x ∈ E ε l , β ( x ) , if x ∈ Ω ∖ ⋃ l E ε l , (164) \widehat{\beta}_{\varepsilon}(x):=
\begin{cases}
\beta_{\varepsilon}^{l}(x), & \text{if }x\in E_{\varepsilon}^{l},\\
\beta(x), & \text{if }x\in\Omega\setminus\bigcup_l E_{\varepsilon}^{l},
\end{cases} \tag*{(164)} β ε ( x ) := { β ε l ( x ) , β ( x ) , if x ∈ E ε l , if x ∈ Ω ∖ ⋃ l E ε l , (164)
and combining it with the functions above
β ε ( x ) : = ψ ε o u t ( x ) β ^ ε ( x ) + ψ ε i n ( x ) Id , (165) \beta_{\varepsilon}(x):=\psi_{\varepsilon}^{\mathrm{out}}(x)\widehat{\beta}_{\varepsilon}(x)+\psi_{\varepsilon}^{\mathrm{in}}(x)\operatorname{Id}, \tag*{(165)} β ε ( x ) := ψ ε out ( x ) β ε ( x ) + ψ ε in ( x ) Id , (165)
which may jump only on ⋃ l F l \bigcup_l F_l ⋃ l F l , with β ε ± = ψ ε o u t β ^ ε ± \beta_{\varepsilon}^{\pm}=\psi_{\varepsilon}^{\mathrm{out}}\widehat{\beta}_{\varepsilon}^{\pm} β ε ± = ψ ε out β ε ± .
We have already shown that v ε ∈ H 1 ( Ω ; [ 0 , 1 ] ) v_{\varepsilon}\in H^1(\Omega;[0,1]) v ε ∈ H 1 ( Ω ; [ 0 , 1 ]) ; from the energy estimate it will follow v ε → 1 v_{\varepsilon}\to1 v ε → 1 in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) . From the construction we also have β ε ∈ SBV ( Ω ; M d × d ) \beta_{\varepsilon}\in\operatorname{SBV}(\Omega;\mathbb{M}^{d\times d}) β ε ∈ SBV ( Ω ; M d × d ) with J β ε ⊂ ⋃ l F l J_{\beta_{\varepsilon}}\subset\bigcup_l F_l J β ε ⊂ ⋃ l F l and β ε + = G l β ε − \beta_{\varepsilon}^{+}=G_l\beta_{\varepsilon}^{-} β ε + = G l β ε − on F l F_l F l ; in order to prove β ε ∈ SBV G ( Ω ; M d × d ) \beta_{\varepsilon}\in\operatorname{SBV}_G(\Omega;\mathbb{M}^{d\times d}) β ε ∈ SBV G ( Ω ; M d × d ) it only remains to show that ∇ β ε ∈ L 1 ( Ω ) \nabla\beta_{\varepsilon}\in L^1(\Omega) ∇ β ε ∈ L 1 ( Ω ) , this will be done in (5.58) below. Finally, we notice that, since ρ ε ≤ M ε \rho_{\varepsilon}\le M\varepsilon ρ ε ≤ M ε , in [ Ω ∖ ⋃ l E ε l ] ∖ ( Γ ) M ε + ρ ε \left[\Omega\setminus\bigcup_l E_{\varepsilon}^{l}\right]\setminus(\Gamma)_{M\varepsilon+\rho_{\varepsilon}} [ Ω ∖ ⋃ l E ε l ] ∖ ( Γ ) M ε + ρ ε we have β ε = β \beta_{\varepsilon}=\beta β ε = β , and therefore ∇ β ε = 0 \nabla\beta_{\varepsilon}=0 ∇ β ε = 0 . As ∥ β ε ∥ L ∞ ≤ d \|\beta_{\varepsilon}\|_{L^\infty}\le\sqrt{d} ∥ β ε ∥ L ∞ ≤ d , we immediately obtain β ε → β \beta_{\varepsilon}\to\beta β ε → β in L 1 ( Ω ; M d × d ) L^1(\Omega;\mathbb{M}^{d\times d}) L 1 ( Ω ; M d × d ) .
We now estimate the different terms of the energy. In ( Γ ) 2 M ε (\Gamma)_{2M\varepsilon} ( Γ ) 2 M ε we have v ε 0 = 0 v_{\varepsilon}^{0}=0 v ε 0 = 0 and therefore f ε ( v ε ) = 0 f_{\varepsilon}(v_{\varepsilon})=0 f ε ( v ε ) = 0 , so that the term f ε 2 ( v ε ) ∣ ∇ β ε ∣ 2 f_{\varepsilon}^{2}(v_{\varepsilon})|\nabla\beta_{\varepsilon}|^2 f ε 2 ( v ε ) ∣∇ β ε ∣ 2 vanishes outside { ψ ε o u t = 1 } ∩ ⋃ l E ε l \{\psi_{\varepsilon}^{\mathrm{out}}=1\}\cap\bigcup_l E_{\varepsilon}^{l} { ψ ε out = 1 } ∩ ⋃ l E ε l . Using (5.43) leads to
F ε ( β ε , v ε ) ≤ ∑ l F ε ( β ε l , v ε l ; E ε l ) + P F ε ( v ε 0 ) + 1 δ ε ∫ ( Γ ) M ε + ρ ε dist 2 ( β ε , O ( d ) ) d x . (166) \mathcal{F}_{\varepsilon}(\beta_{\varepsilon},v_{\varepsilon})\le\sum_l\mathcal{F}_{\varepsilon}(\beta_{\varepsilon}^{l},v_{\varepsilon}^{l};E_{\varepsilon}^{l})+P\mathcal{F}_{\varepsilon}(v_{\varepsilon}^{0})+\frac{1}{\delta_{\varepsilon}}\int_{(\Gamma)_{M\varepsilon+\rho_{\varepsilon}}}\operatorname{dist}^2(\beta_{\varepsilon},O(d))\,dx. \tag*{(166)} F ε ( β ε , v ε ) ≤ l ∑ F ε ( β ε l , v ε l ; E ε l ) + P F ε ( v ε 0 ) + δ ε 1 ∫ ( Γ ) M ε + ρ ε dist 2 ( β ε , O ( d )) d x . (166)
We next estimate these three terms, starting with F ε ( β ε l , v ε l ; E ε l ) \mathcal{F}_{\varepsilon}(\beta_{\varepsilon}^{l},v_{\varepsilon}^{l};E_{\varepsilon}^{l}) F ε ( β ε l , v ε l ; E ε l ) , for a fixed l l l . We denote by λ ^ \widehat{\lambda} λ either λ < + ∞ \lambda<+\infty λ < + ∞ or λ ε \lambda_{\varepsilon} λ ε in the case λ = + ∞ \lambda=+\infty λ = + ∞ . Recalling that ∣ β ε l ∣ ≤ d |\beta_{\varepsilon}^{l}|\le\sqrt{d} ∣ β ε l ∣ ≤ d pointwise, we have
1 δ ε dist 2 ( β ε l ( y + t ν l ) , O ( d ) ) ≤ λ ^ ε dist 2 ( β ε l ( y + t ν l ) , O ( d ) ) + 2 d ε ( ε δ ε − λ ^ ) + . (167) \frac{1}{\delta_{\varepsilon}}\operatorname{dist}^2\left(\beta_{\varepsilon}^{l}(y+t\nu_l),O(d)\right)\le\frac{\widehat{\lambda}}{\varepsilon}\operatorname{dist}^2\left(\beta_{\varepsilon}^{l}(y+t\nu_l),O(d)\right)+\frac{2d}{\varepsilon}\left(\frac{\varepsilon}{\delta_{\varepsilon}}-\widehat{\lambda}\right)_{+}. \tag*{(167)} δ ε 1 dist 2 ( β ε l ( y + t ν l ) , O ( d ) ) ≤ ε λ dist 2 ( β ε l ( y + t ν l ) , O ( d ) ) + ε 2 d ( δ ε ε − λ ) + . (167)
Using (167) , Fubini’s theorem, and changing variables to s : = t / ε s := t/\varepsilon s := t / ε , we obtain
F ε ( β ε l , v ε l ; E ε l ) ≤ ∫ { y ∈ F l : d l ( y ) > c ∗ L } ∫ − L L [ f ε 2 ( v ~ l ) ε ∣ β ~ l ′ ∣ 2 + λ ^ dist 2 ( β ~ l , O ( d ) ) \mathcal{F}_{\varepsilon}(\beta_{\varepsilon}^{l},v_{\varepsilon}^{l};E_{\varepsilon}^{l})\leq\int_{\{y\in F_{l}:d_{l}(y)>c_{*}L\}}\int_{-L}^{L}\left[\frac{f_{\varepsilon}^{2}(\widetilde{v}_{l})}{\varepsilon}|\widetilde{\beta}_{l}'|^{2}+\widehat{\lambda}\operatorname{dist}^{2}(\widetilde{\beta}_{l},O(d))\right. F ε ( β ε l , v ε l ; E ε l ) ≤ ∫ { y ∈ F l : d l ( y ) > c ∗ L } ∫ − L L [ ε f ε 2 ( v l ) ∣ β l ′ ∣ 2 + λ dist 2 ( β l , O ( d ))
+ 2 d ( ε δ ε − λ ^ ) + + ( 1 − v ~ l ) 2 4 + ∣ v ~ l ′ ∣ 2 ] d s d H N − 1 ( y ) . (168) \left.+2d\left(\frac{\varepsilon}{\delta_{\varepsilon}}-\widehat{\lambda}\right)_{+}+\frac{(1-\widetilde{v}_{l})^{2}}{4}+|\widetilde{v}_{l}'|^{2}\right]\,ds\,d\mathcal{H}^{N-1}(y). \tag*{(168)} + 2 d ( δ ε ε − λ ) + + 4 ( 1 − v l ) 2 + ∣ v l ′ ∣ 2 ] d s d H N − 1 ( y ) . (168)
Estimating f ε 2 ( v ~ l ) ≤ ε f 2 ( v ~ l ) f_{\varepsilon}^{2}(\widetilde{v}_{l})\leq\varepsilon f^{2}(\widetilde{v}_{l}) f ε 2 ( v l ) ≤ ε f 2 ( v l ) , and then using (156) and (158) (in the case λ = + ∞ \lambda=+\infty λ = + ∞ ), leads to
F ε ( β ε l , v ε l ; E ε l ) ≤ H N − 1 ( F l ) [ g λ ( R l − , R l + ) + 4 d L ( ε δ ε − λ ^ ) + + ω ( L ) + ω ( λ ε ) ] , (169) \mathcal{F}_{\varepsilon}(\beta_{\varepsilon}^{l},v_{\varepsilon}^{l};E_{\varepsilon}^{l})\leq\mathcal{H}^{N-1}(F_{l})\left[g_{\lambda}(R_{l}^{-},R_{l}^{+})+4dL\left(\frac{\varepsilon}{\delta_{\varepsilon}}-\widehat{\lambda}\right)_{+}+\omega(L)+\omega(\lambda_{\varepsilon})\right], \tag*{(169)} F ε ( β ε l , v ε l ; E ε l ) ≤ H N − 1 ( F l ) [ g λ ( R l − , R l + ) + 4 d L ( δ ε ε − λ ) + + ω ( L ) + ω ( λ ε ) ] , (169)
with the term ω ( λ ε ) \omega(\lambda_{\varepsilon}) ω ( λ ε ) only present if λ = + ∞ \lambda=+\infty λ = + ∞ . For later reference, we note that a similar but simpler computation, using again (156) and (158) , leads to
∥ ∇ β ε ∥ L 2 ( E ε l ) 2 ≤ 1 ε H N − 1 ( F l ) ∥ β ~ l ′ ∥ L 2 ( R ) 2 ≤ C ( L ) 1 + λ ^ ε . (170) \|\nabla\beta_{\varepsilon}\|_{L^{2}(E_{\varepsilon}^{l})}^{2}\leq\frac{1}{\varepsilon}\mathcal{H}^{N-1}(F_{l})\|\widetilde{\beta}_{l}'\|_{L^{2}(\mathbb{R})}^{2}\leq C(L)\frac{1+\widehat{\lambda}}{\varepsilon}. \tag*{(170)} ∥∇ β ε ∥ L 2 ( E ε l ) 2 ≤ ε 1 H N − 1 ( F l ) ∥ β l ′ ∥ L 2 ( R ) 2 ≤ C ( L ) ε 1 + λ . (170)
The second term in (166) was already treated in (155) . In the third term we use the bound ∥ β ε ∥ L ∞ ≤ d \|\beta_{\varepsilon}\|_{L^{\infty}}\leq\sqrt{d} ∥ β ε ∥ L ∞ ≤ d and the fact that Γ \Gamma Γ is a finite union of ( N − 2 ) (N-2) ( N − 2 ) -dimensional polyhedra to obtain
1 δ ε ∫ ( Γ ) M ε + ρ ε dist 2 ( β ε , O ( d ) ) d x ≤ C δ ε ∣ ( Γ ) M ε + ρ ε ∖ ( Γ ) M ε − ρ ε ∣ ≤ C ρ ε L ε δ ε . (171) \frac{1}{\delta_{\varepsilon}}\int_{(\Gamma)_{M_{\varepsilon}+\rho_{\varepsilon}}}\operatorname{dist}^{2}(\beta_{\varepsilon},O(d))\,dx\leq\frac{C}{\delta_{\varepsilon}}|(\Gamma)_{M_{\varepsilon}+\rho_{\varepsilon}}\setminus(\Gamma)_{M_{\varepsilon}-\rho_{\varepsilon}}|\leq C\frac{\rho_{\varepsilon}L\varepsilon}{\delta_{\varepsilon}}. \tag*{(171)} δ ε 1 ∫ ( Γ ) M ε + ρ ε dist 2 ( β ε , O ( d )) d x ≤ δ ε C ∣ ( Γ ) M ε + ρ ε ∖ ( Γ ) M ε − ρ ε ∣ ≤ C δ ε ρ ε L ε . (171)
We are now ready to conclude the estimate of the energy.
We first consider the case λ < + ∞ \lambda<+\infty λ < + ∞ , taking the limit as ε → 0 \varepsilon\to0 ε → 0 first with L L L fixed. By (171) if we choose ρ ε \rho_{\varepsilon} ρ ε such that
lim ε → 0 ρ ε ε δ ε = 0 , (172) \lim_{\varepsilon\to0}\frac{\rho_{\varepsilon}\varepsilon}{\delta_{\varepsilon}}=0, \tag*{(172)} ε → 0 lim δ ε ρ ε ε = 0 , (172)
then
lim sup ε → 0 1 δ ε ∫ Ω ∖ ⋃ l E ε l dist 2 ( β ε , O ( d ) ) d x = 0. (173) \limsup_{\varepsilon\to0}\frac{1}{\delta_{\varepsilon}}\int_{\Omega\setminus\bigcup_{l}E_{\varepsilon}^{l}}\operatorname{dist}^{2}(\beta_{\varepsilon},O(d))\,dx=0. \tag*{(173)} ε → 0 lim sup δ ε 1 ∫ Ω ∖ ⋃ l E ε l dist 2 ( β ε , O ( d )) d x = 0. (173)
Starting from (166) , taking the limit as ε → 0 \varepsilon\to0 ε → 0 , using (169) , (155) , and (173) leads to
lim sup ε → 0 F ε ( β ε , v ε ) ≤ ∑ l H N − 1 ( F l ) ( g λ ( R l − , R l + ) + ω ( L ) ) , \limsup_{\varepsilon\to0}\mathcal{F}_{\varepsilon}(\beta_{\varepsilon},v_{\varepsilon})\leq\sum_{l}\mathcal{H}^{N-1}(F_{l})\bigl(g_{\lambda}(R_{l}^{-},R_{l}^{+})+\omega(L)\bigr), ε → 0 lim sup F ε ( β ε , v ε ) ≤ l ∑ H N − 1 ( F l ) ( g λ ( R l − , R l + ) + ω ( L ) ) ,
so that
lim sup L → ∞ lim sup ε → 0 F ε ( β ε , v ε ) ≤ ∑ l H N − 1 ( F l ) g λ ( R l − , R l + ) = F λ ( β , 1 ) , \limsup_{L\to\infty}\limsup_{\varepsilon\to0}\mathcal{F}_{\varepsilon}(\beta_{\varepsilon},v_{\varepsilon})\leq\sum_{l}\mathcal{H}^{N-1}(F_{l})g_{\lambda}(R_{l}^{-},R_{l}^{+})=\mathcal{F}^{\lambda}(\beta,1), L → ∞ lim sup ε → 0 lim sup F ε ( β ε , v ε ) ≤ l ∑ H N − 1 ( F l ) g λ ( R l − , R l + ) = F λ ( β , 1 ) ,
and the conclusion follows from a diagonal argument.
It remains to address the L 2 L^{2} L 2 estimate on ∇ β ε \nabla\beta_{\varepsilon} ∇ β ε . We recall the definitions in (165) and (163) , as well as the bound ∣ β ^ ε ∣ ≤ d |\widehat{\beta}_{\varepsilon}|\leq\sqrt{d} ∣ β ε ∣ ≤ d , and write the pointwise estimate
∣ ∇ β ε ∣ ≤ d ( ∣ ∇ ψ ε o u t ∣ + ∣ ∇ ψ ε i n ∣ ) + ∣ ψ ε o u t ∇ β ^ ε ∣ . (174) |\nabla\beta_{\varepsilon}|\leq\sqrt{d}\bigl(|\nabla\psi_{\varepsilon}^{\mathrm{out}}|+|\nabla\psi_{\varepsilon}^{\mathrm{in}}|\bigr)+|\psi_{\varepsilon}^{\mathrm{out}}\nabla\widehat{\beta}_{\varepsilon}|. \tag*{(174)} ∣∇ β ε ∣ ≤ d ( ∣∇ ψ ε out ∣ + ∣∇ ψ ε in ∣ ) + ∣ ψ ε out ∇ β ε ∣. (174)
We treat the two terms separately. First,
∥ ∇ ψ ε o u t ∥ L 2 ( Ω ) 2 + ∥ ∇ ψ ε i n ∥ L 2 ( Ω ) 2 ≤ C ρ ε 2 ∣ ( Γ ) M ε + ρ ε ∖ ( Γ ) M ε − ρ ε ∣ ≤ C M ε ρ ε , (175) \|\nabla\psi_{\varepsilon}^{\mathrm{out}}\|_{L^{2}(\Omega)}^{2}+\|\nabla\psi_{\varepsilon}^{\mathrm{in}}\|_{L^{2}(\Omega)}^{2}\leq\frac{C}{\rho_{\varepsilon}^{2}}|(\Gamma)_{M_{\varepsilon}+\rho_{\varepsilon}}\setminus(\Gamma)_{M_{\varepsilon}-\rho_{\varepsilon}}|\leq C\frac{M\varepsilon}{\rho_{\varepsilon}}, \tag*{(175)} ∥∇ ψ ε out ∥ L 2 ( Ω ) 2 + ∥∇ ψ ε in ∥ L 2 ( Ω ) 2 ≤ ρ ε 2 C ∣ ( Γ ) M ε + ρ ε ∖ ( Γ ) M ε − ρ ε ∣ ≤ C ρ ε M ε , (175)
where we used that Γ \Gamma Γ is a finite union of N − 2 N-2 N − 2 dimensional polyhedra. For the second one, from (164) and ∇ β = 0 \nabla\beta=0 ∇ β = 0 we see that we can restrict to ⋃ l E ε l \bigcup_{l}E_{\varepsilon}^{l} ⋃ l E ε l , that were already treated. Recalling (174) , (170) , and (175) ,
∥ ∇ β ε ∥ L 2 ( Ω ) 2 ≤ C ( L ε ρ ε + C ( L ) 1 + λ ^ ε ) . (176) \|\nabla\beta_{\varepsilon}\|_{L^{2}(\Omega)}^{2}\leq C\left(\frac{L\varepsilon}{\rho_{\varepsilon}}+C(L)\frac{1+\widehat{\lambda}}{\varepsilon}\right). \tag*{(176)} ∥∇ β ε ∥ L 2 ( Ω ) 2 ≤ C ( ρ ε L ε + C ( L ) ε 1 + λ ) . (176)
It remains to choose ρ ε \rho_\varepsilon ρ ε . We recall (172) and see that we need to choose ρ ε → 0 \rho_\varepsilon\to0 ρ ε → 0 , 0 < ρ ε ≤ ε 0 < \rho_\varepsilon\le\varepsilon 0 < ρ ε ≤ ε , such that
lim ε → 0 ρ ε ε δ ε = 0 and lim ε → 0 η ε [ 1 ε + ε ρ ε ] = 0. (177) \lim_{\varepsilon\to0} \frac{\rho_\varepsilon\varepsilon}{\delta_\varepsilon}=0 \quad\text{and}\quad\lim_{\varepsilon\to0}\eta_\varepsilon\left[\frac{1}{\varepsilon}+\frac{\varepsilon}{\rho_\varepsilon}\right]=0. \tag*{(177)} ε → 0 lim δ ε ρ ε ε = 0 and ε → 0 lim η ε [ ε 1 + ρ ε ε ] = 0. (177)
We fix ρ ε : = ε \rho_\varepsilon:= \varepsilon ρ ε := ε and see that it obeys all requirements provided that
lim ε → 0 η ε ε = 0 , (178) \lim_{\varepsilon\to0}\frac{\eta_\varepsilon}{\varepsilon}=0, \tag*{(178)} ε → 0 lim ε η ε = 0 , (178)
as in the assumptions.
With this choice of η ε \eta_\varepsilon η ε we then obtain that
lim sup L → + ∞ lim sup ε → 0 η ε ∥ ∇ β ε ∥ L 2 ( Ω ) 2 = 0 (179) \limsup_{L\to+\infty}\limsup_{\varepsilon\to0}\eta_\varepsilon\lVert\nabla\beta_\varepsilon\rVert_{L^2(\Omega)}^2=0 \tag*{(179)} L → + ∞ lim sup ε → 0 lim sup η ε ∥ ∇ β ε ∥ L 2 ( Ω ) 2 = 0 (179)
so that with a diagonal argument we conclude the proof in the case λ < + ∞ \lambda<+\infty λ < + ∞ .
We now turn to the case λ = + ∞ \lambda=+\infty λ = + ∞ . We fix a sequence L ε → + ∞ L_\varepsilon\to+\infty L ε → + ∞ chosen below, that will depend on the sequence λ ε \lambda_\varepsilon λ ε defined in (157) .
Now starting from (166) , collecting (169) , (155) , (171) leads to
F ε ( β ε , v ε ) ≤ ∑ l H N − 1 ( F l ) [ g ∞ ( R l − , R l + ) + ω ε ] + C L ε 2 + C ρ ε L ε ε δ ε (180) \mathcal{F}_\varepsilon(\beta_\varepsilon,v_\varepsilon)\leq\sum_l\mathcal{H}^{N-1}(F_l)\left[g_\infty(R_l^-,R_l^+)+\omega_\varepsilon\right]+CL_\varepsilon^2+C\frac{\rho_\varepsilon L_\varepsilon\varepsilon}{\delta_\varepsilon} \tag*{(180)} F ε ( β ε , v ε ) ≤ l ∑ H N − 1 ( F l ) [ g ∞ ( R l − , R l + ) + ω ε ] + C L ε 2 + C δ ε ρ ε L ε ε (180)
where ω ε = ω ( L ε ) + ω ( λ ε ) \omega_\varepsilon=\omega(L_\varepsilon)+\omega(\lambda_\varepsilon) ω ε = ω ( L ε ) + ω ( λ ε ) . Then choosing L ε → ∞ L_\varepsilon\to\infty L ε → ∞ so that L ε 2 → 0 L_\varepsilon^2\to0 L ε 2 → 0 as ε → 0 \varepsilon\to0 ε → 0 , and recalling (172) , we conclude that
lim sup ε → 0 F ε ( β ε , v ε ) ≤ ∑ l H N − 1 ( F l ) g ∞ ( R l − , R l + ) = F ∞ ( β , 1 ) . \limsup_{\varepsilon\to0}\mathcal{F}_\varepsilon(\beta_\varepsilon,v_\varepsilon)\leq\sum_l\mathcal{H}^{N-1}(F_l)g_\infty(R_l^-,R_l^+)=\mathcal{F}^\infty(\beta,1). ε → 0 lim sup F ε ( β ε , v ε ) ≤ l ∑ H N − 1 ( F l ) g ∞ ( R l − , R l + ) = F ∞ ( β , 1 ) .
In order to obtain the estimate of the gradients, in view of (176) it is enough to replace the conditions (177) with
lim ε → 0 ρ ε L ε ε δ ε = 0 and lim ε → 0 η ε [ C ( L ε ) λ ε ε + L ε ε ρ ε ] = 0. (181) \lim_{\varepsilon\to0}\frac{\rho_\varepsilon L_\varepsilon\varepsilon}{\delta_\varepsilon}=0\quad\text{and}\quad\lim_{\varepsilon\to0}\eta_\varepsilon\left[C(L_\varepsilon)\frac{\lambda_\varepsilon}{\varepsilon}+\frac{L_\varepsilon\varepsilon}{\rho_\varepsilon}\right]=0. \tag*{(181)} ε → 0 lim δ ε ρ ε L ε ε = 0 and ε → 0 lim η ε [ C ( L ε ) ε λ ε + ρ ε L ε ε ] = 0. (181)
We fix ρ ε : = δ ε ∧ ε \rho_\varepsilon:=\delta_\varepsilon\wedge\varepsilon ρ ε := δ ε ∧ ε , so that the first condition in (181) holds. We can choose L ε → ∞ L_\varepsilon\to\infty L ε → ∞ such that the second one holds if and only if
lim ε → 0 η ε λ ε ε = lim ε → 0 η ε δ ε = 0 , and lim ε → 0 η ε ε ρ ε = 0. (182) \lim_{\varepsilon\to0}\eta_\varepsilon\frac{\lambda_\varepsilon}{\varepsilon}=\lim_{\varepsilon\to0}\frac{\eta_\varepsilon}{\delta_\varepsilon}=0,\quad\text{and}\quad\lim_{\varepsilon\to0}\eta_\varepsilon\frac{\varepsilon}{\rho_\varepsilon}=0. \tag*{(182)} ε → 0 lim η ε ε λ ε = ε → 0 lim δ ε η ε = 0 , and ε → 0 lim η ε ρ ε ε = 0. (182)
Remark 5.7. The upper bound (UB)’ in Theorem 2.15 follows directly from Theorem 5.6 and Theorem 2.12 . Namely, given u ∈ S B V 0 ( Ω ; O ( d ) / G ) u\in\mathrm{SBV}_0(\Omega;\mathrm{O}(d)/G) u ∈ SBV 0 ( Ω ; O ( d ) / G ) , by Remark 2.1, (51) , and (52) , a recovery sequence for F ε λ ( u , 1 ) \mathcal{F}^{\lambda}_\varepsilon(u,1) F ε λ ( u , 1 ) is given by ( β ε , v ε ) ε (\beta_\varepsilon,v_\varepsilon)_\varepsilon ( β ε , v ε ) ε , for ( β ε , v ε ) ε (\beta_\varepsilon,v_\varepsilon)_\varepsilon ( β ε , v ε ) ε recovery sequence for F λ ( β , 1 ) \mathcal{F}^{\lambda}(\beta,1) F λ ( β , 1 ) , β ∈ S B V 0 ( Ω ; O ( d ) ) \beta\in\mathrm{SBV}_0(\Omega;\mathrm{O}(d)) β ∈ SBV 0 ( Ω ; O ( d )) being the lifting of u u u given in Theorem 2.12 (see also Remark 2.13).
Therefore the proofs of Theorems 1.1 and 2.15 are completed.
Acknowledgements
We thank Luigi Ambrosio for an interesting comment on a preliminary version of this work. S.C. gratefully thanks Dipartimento di Matematica Guido Castelnuovo, Sapienza Università di Roma, where a large part of this work was carried out, for the warm hospitality. This work was partially funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), project CRC 1720 – 539309657. V.C., A.G., A.M. acknowledge the support of the MUR - PRIN project 2022J4FYNJ CUP B53D23009320006 “Variational methods for stationary and evolution problems with singularities and interfaces”, PNRR Italia Domani, funded by the European Union via the program NextGenerationEU. V.C. acknowledges also the support of the SEED-PNR Project CUP B87G22001200001. V.C., A.G., A.M. are members of the Gruppo Nazionale per l’Analisi Matematica, la Probabilità e le loro Applicazioni (INdAM-GNAMPA).
References
Sergio Conti: Institut für Angewandte Mathematik, Universität Bonn, 53115 Bonn, Germany, Vito Crismale, Adriana Garroni, Annalisa Malusa: Dipartimento di Matematica, Università La Sapienza di Roma, piazzale Aldo Moro 5, 00185 Roma